From c882af412c45faa9870f26a1f5fb0e65797d9dbc Mon Sep 17 00:00:00 2001
From: Illia Oleksiienko <io@ece.au.dk>
Date: Tue, 18 Oct 2022 22:58:47 +0000
Subject: [PATCH] Add tapp, uncertainty outputs
---
run/eval_tracking_3d.py | 225 +++++++++++++
run/train_3d.py | 72 +++-
src/opendr/engine/target.py | 307 +++++++++++++++++-
.../configs/tanet/car/tapp_16.proto | 173 ++++++++++
.../configs/tanet/car/vnn_tapp_16.proto | 173 ++++++++++
.../second_detector/data/kitti_common.py | 14 +-
.../pytorch/core/box_coders.py | 4 +
.../pytorch/core/box_torch_ops.py | 84 +++++
.../second_detector/pytorch/models/tanet.py | 18 +-
.../pytorch/models/voxelnet.py | 209 ++++++++++--
.../second_detector/run.py | 28 +-
.../voxel_object_detection_3d_learner.py | 68 +++-
12 files changed, 1312 insertions(+), 63 deletions(-)
create mode 100644 run/eval_tracking_3d.py
create mode 100644 src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/tapp_16.proto
create mode 100644 src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/vnn_tapp_16.proto
diff --git a/run/eval_tracking_3d.py b/run/eval_tracking_3d.py
new file mode 100644
index 0000000000..3275444c00
--- /dev/null
+++ b/run/eval_tracking_3d.py
@@ -0,0 +1,225 @@
+import sys
+import os
+import torch
+import tqdm
+import fire
+from opendr.engine.target import BoundingBox3D
+from opendr.engine.datasets import PointCloudsDatasetIterator
+from opendr.perception.object_detection_3d import VoxelObjectDetection3DLearner
+from opendr.perception.object_detection_3d import (
+ LabeledPointCloudsDatasetIterator,
+)
+from opendr.perception.object_tracking_3d.datasets.kitti_tracking import LabeledTrackingPointCloudsDatasetIterator
+
+config_roots = {
+ "pointpillars": os.path.join(
+ ".",
+ "src",
+ "opendr",
+ "perception",
+ "object_detection_3d",
+ "voxel_object_detection_3d",
+ "second_detector",
+ "configs",
+ "pointpillars",
+ "car",
+ ),
+ "tanet": os.path.join(
+ ".",
+ "src",
+ "opendr",
+ "perception",
+ "object_detection_3d",
+ "voxel_object_detection_3d",
+ "second_detector",
+ "configs",
+ "tanet",
+ "car",
+ ),
+}
+temp_dir = "./run/models"
+subsets_path = os.path.join(
+ ".",
+ "src",
+ "opendr",
+ "perception",
+ "object_detection_3d",
+ "datasets",
+ "kitti_subsets",
+)
+
+dataset_tracking_path = "/data/sets/kitti_tracking"
+datasets = {}
+
+all_track_ids = ["0000"]
+
+for track_id in all_track_ids:
+ datasets[track_id] = LabeledTrackingPointCloudsDatasetIterator(
+ dataset_tracking_path + "/training/velodyne/" + track_id,
+ dataset_tracking_path + "/training/label_02/" + track_id + ".txt",
+ dataset_tracking_path + "/training/calib/" + track_id + ".txt",
+ )
+
+
+def save_detection_inference(
+ model_type,
+ device="cuda:0",
+ name="pointpillars_car",
+ samples_list=[1],
+ config="xyres_16.proto",
+ eval_suffix="classic",
+):
+
+ config = os.path.join(config_roots[model_type], config,)
+ model_path = os.path.join(temp_dir, name)
+
+ learner = VoxelObjectDetection3DLearner(
+ model_config_path=config, device=device, checkpoint_after_iter=1000, return_uncertainty=True
+ )
+ learner.load(model_path)
+
+ results = {}
+ for samples in samples_list:
+ print("samples =", samples)
+
+ for track_id in all_track_ids:
+
+ dataset = datasets[track_id]
+
+ os.makedirs(os.path.join(
+ model_path, "tracking_inference_detections", eval_suffix, "samples_" + str(samples)
+ ), exist_ok=True)
+
+ with open(
+ os.path.join(
+ model_path,
+ "tracking_inference_detections",
+ eval_suffix,
+ "samples_" + str(samples),
+ track_id + ".txt"
+ ), "w"
+ ) as f:
+ for frame, (input, _) in tqdm.tqdm(enumerate(dataset), total=len(dataset)):
+ output = learner.infer(input, samples=samples)
+
+ result = "\n".join(box.to_kitti_tracking_string(frame) for box in output)
+
+ if len(output) > 0:
+ result += "\n"
+
+ f.write(result)
+
+
+
+def test_model(
+ model_type,
+ device="cuda:0",
+ name="pointpillars_car",
+ samples_list=[1],
+ config="xyres_16.proto",
+ eval_suffix="classic",
+):
+
+ return save_detection_inference(
+ model_type=model_type,
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+ # config = os.path.join(config_roots[model_type], config,)
+ # model_path = os.path.join(temp_dir, name)
+
+ # learner = VoxelObjectDetection3DLearner(
+ # model_config_path=config, device=device, checkpoint_after_iter=1000
+ # )
+ # learner.load(model_path)
+
+
+
+def test_pointpillars(
+ device="cuda:0",
+ name="pointpillars_car",
+ samples_list=[1],
+ config="xyres_16.proto",
+ eval_suffix="classic",
+):
+ return test_model(
+ "pointpillars",
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
+def test_tanet(
+ device="cuda:0",
+ name="tanet_car",
+ samples_list=[1],
+ config="xyres_16.proto",
+ eval_suffix="classic",
+):
+ return test_model(
+ "tanet",
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
+def test_vnn_pointpillars(
+ device="cuda:0",
+ name="vnn_pointpillars_car",
+ samples_list=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
+ config="vnn_xyres_16.proto",
+ eval_suffix="vnn",
+):
+ return test_pointpillars(
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+def test_vnna_pointpillars(
+ device="cuda:0",
+ name="vnna_pointpillars_car",
+ samples_list=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
+ config="vnna_xyres_16.proto",
+ eval_suffix="vnn",
+):
+ return test_pointpillars(
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
+def test_vnn_tanet(
+ device="cuda:0",
+ name="vnn_tanet_car",
+ samples_list=[1, 2, 3, 4],
+ config="vnn_xyres_16.proto",
+ eval_suffix="vnn",
+):
+ return test_model(
+ "tanet",
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
+if __name__ == "__main__":
+ fire.Fire()
diff --git a/run/train_3d.py b/run/train_3d.py
index 0d5c4f5b82..4be41991e3 100644
--- a/run/train_3d.py
+++ b/run/train_3d.py
@@ -64,6 +64,7 @@ def save_model(
device=device,
checkpoint_after_iter=1000,
checkpoint_load_iter=load,
+ return_uncertainty=True
)
learner.save(model_path)
@@ -95,6 +96,7 @@ def train_model(
device=device,
checkpoint_after_iter=1000,
checkpoint_load_iter=load,
+ return_uncertainty=True
)
learner.fit(
@@ -137,6 +139,23 @@ def train_tanet(
)
+def train_tapp(
+ device="cuda:0",
+ load=0,
+ name="tapp_car",
+ config="tapp_16.proto",
+ samples_list=None,
+):
+ return train_model(
+ "tanet",
+ device=device,
+ load=load,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ )
+
+
def train_vnn_pointpillars(
device="cuda:0",
load=0,
@@ -178,6 +197,23 @@ def train_vnn_tanet(
)
+def train_vnn_tapp(
+ device="cuda:0",
+ load=0,
+ name="vnn_tapp_car",
+ config="vnn_tapp_16.proto",
+ samples_list=[2, 1],
+):
+ return train_model(
+ "tanet",
+ device=device,
+ load=load,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ )
+
+
def test_model(
model_type,
device="cuda:0",
@@ -191,7 +227,7 @@ def test_model(
model_path = os.path.join(temp_dir, name)
learner = VoxelObjectDetection3DLearner(
- model_config_path=config, device=device, checkpoint_after_iter=1000
+ model_config_path=config, device=device, checkpoint_after_iter=1000, return_uncertainty=True
)
learner.load(model_path)
@@ -246,6 +282,23 @@ def test_tanet(
)
+def test_tapp(
+ device="cuda:0",
+ name="tapp_car",
+ samples_list=[1],
+ config="tapp_16.proto",
+ eval_suffix="classic",
+):
+ return test_model(
+ "tanet",
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
def test_vnn_pointpillars(
device="cuda:0",
name="vnn_pointpillars_car",
@@ -294,5 +347,22 @@ def test_vnn_tanet(
)
+def test_vnn_tapp(
+ device="cuda:0",
+ name="vnn_tapp_car",
+ samples_list=[1, 2, 3, 4],
+ config="vnn_tapp_16.proto",
+ eval_suffix="vnn",
+):
+ return test_model(
+ "tanet",
+ device=device,
+ name=name,
+ samples_list=samples_list,
+ config=config,
+ eval_suffix=eval_suffix,
+ )
+
+
if __name__ == "__main__":
fire.Fire()
diff --git a/src/opendr/engine/target.py b/src/opendr/engine/target.py
index 6b26fca636..95b1937ccb 100644
--- a/src/opendr/engine/target.py
+++ b/src/opendr/engine/target.py
@@ -675,18 +675,18 @@ class BoundingBox3D(Target):
return result
def to_kitti_tracking_string(self, frame):
- result = " ".join([
+ result = " ".join(map(str, [
frame,
self.name,
- self.truncated,
- self.occluded,
- self.alpha,
- *self.bbox2d, # x y w h
+ float(self.truncated),
+ float(self.occluded),
+ float(self.alpha),
+ *([0,0,0,0] if self.bbox2d is None else self.bbox2d), # x y w h
*self.dimensions, # w h l
*self.location, # x y z
self.rotation_y,
- self.score,
- ])
+ float(self.confidence),
+ ]))
return result
@@ -1111,3 +1111,296 @@ class Heatmap(Target):
:rtype: str
"""
return str(self.data)
+
+
+
+class UncertaintyBoundingBox3D(BoundingBox3D):
+ """
+ This target is used for 3D Object Detection and Tracking with uncertainty.
+ A bounding box is described by its location (x, y, z), dimensions (w, h, d) and rotation (along vertical y axis).
+ Additional fields are used to describe confidence (score), 2D projection of the box on camera image (bbox2d),
+ truncation (truncated) and occlusion (occluded) levels, the name of an object (name) and
+ observation angle of an object (alpha).
+ UncertaintyBoundingBox3D provides uncertainty values for all regressed fields (this excludes name).
+ """
+
+ def __init__(
+ self,
+ name,
+ truncated,
+ occluded,
+ alpha,
+ bbox2d,
+ dimensions,
+ location,
+ rotation_y,
+ variance_truncated,
+ variance_occluded,
+ variance_alpha,
+ variance_bbox2d,
+ variance_dimensions,
+ variance_location,
+ variance_rotation_y,
+ score=0,
+ variance_score=0,
+ ):
+ super().__init__(
+ name,
+ truncated,
+ occluded,
+ alpha,
+ bbox2d,
+ dimensions,
+ location,
+ rotation_y,
+ score=0,
+ )
+ self.data.update({
+ "variance_truncated": variance_truncated,
+ "variance_occluded": variance_occluded,
+ "variance_alpha": variance_alpha,
+ "variance_bbox2d": variance_bbox2d,
+ "variance_dimensions": variance_dimensions,
+ "variance_location": variance_location,
+ "variance_rotation_y": variance_rotation_y,
+ })
+ self.confidence = score
+ self.variance_confidence = variance_score
+
+ def kitti(self):
+ result = super().kitti()
+
+ result["variance_truncated"] = np.array([self.data["variance_truncated"]])
+ result["variance_occluded"] = np.array([self.data["variance_occluded"]])
+ result["variance_alpha"] = np.array([self.data["variance_alpha"]])
+ result["variance_bbox"] = np.array([self.data["variance_bbox2d"]])
+ result["variance_dimensions"] = np.array([self.data["variance_dimensions"]])
+ result["variance_location"] = np.array([self.data["variance_location"]])
+ result["variance_rotation_y"] = np.array([self.data["variance_rotation_y"]])
+ result["variance_score"] = np.array([self.variance_confidence])
+
+ return result
+
+ def to_kitti_tracking_string(self, frame):
+
+ result = super().to_kitti_tracking_string(frame)
+
+ result += " ".join(map(str, [
+ float(self.variance_truncated),
+ float(self.variance_occluded),
+ float(self.variance_alpha),
+ *([0,0,0,0] if self.variance_bbox2d is None else self.variance_bbox2d), # x y w h
+ *self.variance_dimensions, # w h l
+ *self.variance_location, # x y z
+ self.variance_rotation_y,
+ float(self.variance_confidence),
+ ]))
+
+ return result
+
+ @property
+ def name(self):
+ return self.data["name"]
+
+ @property
+ def truncated(self):
+ return self.data["truncated"]
+
+ @property
+ def occluded(self):
+ return self.data["occluded"]
+
+ @property
+ def alpha(self):
+ return self.data["alpha"]
+
+ @property
+ def bbox2d(self):
+ return self.data["bbox2d"]
+
+ @property
+ def dimensions(self):
+ return self.data["dimensions"]
+
+ @property
+ def location(self):
+ return self.data["location"]
+
+ @property
+ def rotation_y(self):
+ return self.data["rotation_y"]
+
+ @property
+ def variance_truncated(self):
+ return self.data["variance_truncated"]
+
+ @property
+ def variance_occluded(self):
+ return self.data["variance_occluded"]
+
+ @property
+ def variance_alpha(self):
+ return self.data["variance_alpha"]
+
+ @property
+ def variance_bbox2d(self):
+ return self.data["variance_bbox2d"]
+
+ @property
+ def variance_dimensions(self):
+ return self.data["variance_dimensions"]
+
+ @property
+ def variance_location(self):
+ return self.data["variance_location"]
+
+ @property
+ def variance_rotation_y(self):
+ return self.data["variance_rotation_y"]
+
+ def __repr__(self):
+ return "UncertaintyBoundingBox3D " + str(self)
+
+ def __str__(self):
+ return str(self.kitti())
+
+
+class UncertaintyBoundingBox3DList(Target):
+ """
+ This target is used for 3D Object Detection with uncertainty. It contains a list of UncertaintyBoundingBox3D targets.
+ A bounding box is described by its location (x, y, z), dimensions (l, h, w) and rotation (along vertical (y) axis).
+ Additional fields are used to describe confidence (score), 2D projection of the box on camera image (bbox2d),
+ truncation (truncated) and occlusion (occluded) levels, the name of an object (name) and
+ observation angle of an object (alpha).
+ UncertaintyBoundingBox3D provides uncertainty values for all regressed fields (this excludes name).
+ """
+
+ def __init__(
+ self,
+ bounding_boxes_3d
+ ):
+ super().__init__()
+ self.data = bounding_boxes_3d
+ self.confidence = None if len(self.data) == 0 else np.mean([box.confidence for box in self.data])
+ self.variance_confidence = None if len(self.data) == 0 else np.mean([box.variance_confidence for box in self.data])
+
+ @staticmethod
+ def from_kitti(boxes_kitti):
+
+ count = len(boxes_kitti["name"])
+
+ boxes3d = []
+
+ for i in range(count):
+ box3d = UncertaintyBoundingBox3D(
+ boxes_kitti["name"][i],
+ boxes_kitti["truncated"][i],
+ boxes_kitti["occluded"][i],
+ boxes_kitti["alpha"][i],
+ boxes_kitti["bbox"][i],
+ boxes_kitti["dimensions"][i],
+ boxes_kitti["location"][i],
+ boxes_kitti["rotation_y"][i],
+ boxes_kitti["variance_truncated"][i],
+ boxes_kitti["variance_occluded"][i],
+ boxes_kitti["variance_alpha"][i],
+ boxes_kitti["variance_bbox"][i],
+ boxes_kitti["variance_dimensions"][i],
+ boxes_kitti["variance_location"][i],
+ boxes_kitti["variance_rotation_y"][i],
+ boxes_kitti["score"][i],
+ boxes_kitti["variance_score"][i],
+ )
+
+ boxes3d.append(box3d)
+
+ return UncertaintyBoundingBox3DList(boxes3d)
+
+ def kitti(self):
+
+ result = {
+ "name": [],
+ "truncated": [],
+ "occluded": [],
+ "alpha": [],
+ "bbox": [],
+ "dimensions": [],
+ "location": [],
+ "rotation_y": [],
+ "variance_truncated": [],
+ "variance_occluded": [],
+ "variance_alpha": [],
+ "variance_bbox": [],
+ "variance_dimensions": [],
+ "variance_location": [],
+ "variance_rotation_y": [],
+ "score": [],
+ "variance_score": [],
+ }
+
+ if len(self.data) == 0:
+ return result
+ elif len(self.data) == 1:
+ return self.data[0].kitti()
+ else:
+
+ for box in self.data:
+ result["name"].append(box.data["name"])
+ result["truncated"].append(box.data["truncated"])
+ result["occluded"].append(box.data["occluded"])
+ result["alpha"].append(box.data["alpha"])
+ result["bbox"].append(box.data["bbox2d"])
+ result["dimensions"].append(box.data["dimensions"])
+ result["location"].append(box.data["location"])
+ result["rotation_y"].append(box.data["rotation_y"])
+ result["variance_truncated"].append(box.data["variance_truncated"])
+ result["variance_occluded"].append(box.data["variance_occluded"])
+ result["variance_alpha"].append(box.data["variance_alpha"])
+ result["variance_bbox"].append(box.data["variance_bbox2d"])
+ result["variance_dimensions"].append(box.data["variance_dimensions"])
+ result["variance_location"].append(box.data["variance_location"])
+ result["variance_rotation_y"].append(box.data["variance_rotation_y"])
+ result["score"].append(box.confidence)
+ result["variance_score"].append(box.variance_confidence)
+
+ result["name"] = np.array(result["name"])
+ result["truncated"] = np.array(result["truncated"])
+ result["occluded"] = np.array(result["occluded"])
+ result["alpha"] = np.array(result["alpha"])
+ result["bbox"] = np.array(result["bbox"])
+ result["dimensions"] = np.array(result["dimensions"])
+ result["location"] = np.array(result["location"])
+ result["rotation_y"] = np.array(result["rotation_y"])
+ result["variance_truncated"] = np.array(result["truncated"])
+ result["variance_occluded"] = np.array(result["occluded"])
+ result["variance_alpha"] = np.array(result["alpha"])
+ result["variance_bbox"] = np.array(result["bbox"])
+ result["variance_dimensions"] = np.array(result["dimensions"])
+ result["variance_location"] = np.array(result["location"])
+ result["variance_rotation_y"] = np.array(result["rotation_y"])
+ result["score"] = np.array(result["score"])
+
+ num_ground_truths = len(result["name"])
+ num_objects = len([x for x in result["name"] if x != "DontCare"])
+ index = list(range(num_objects)) + [-1] * (num_ground_truths - num_objects)
+ result["index"] = np.array(index, dtype=np.int32)
+ result["group_ids"] = np.arange(num_ground_truths, dtype=np.int32)
+
+ return result
+
+ @property
+ def boxes(self):
+ return self.data
+
+ def __getitem__(self, idx):
+ return self.boxes[idx]
+
+ def __len__(self):
+ return len(self.data)
+
+ def __repr__(self):
+ return "UncertaintyBoundingBox3DList " + str(self)
+
+ def __str__(self):
+ return str(self.kitti())
+
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/tapp_16.proto b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/tapp_16.proto
new file mode 100644
index 0000000000..c947358ccb
--- /dev/null
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/tapp_16.proto
@@ -0,0 +1,173 @@
+model: {
+ second: {
+ voxel_generator {
+ point_cloud_range : [0, -39.68, -3, 69.12, 39.68, 1]
+ voxel_size : [0.16, 0.16, 4]
+ max_number_of_points_per_voxel : 100
+ }
+ num_class: 1
+ voxel_feature_extractor: {
+ module_class_name: "PillarFeature_TANet"
+ num_filters: [64]
+ with_distance: false
+ }
+ middle_feature_extractor: {
+ module_class_name: "PointPillarsScatter"
+ }
+ rpn: {
+ module_class_name: "RPN"
+ layer_nums: [3, 5, 5]
+ layer_strides: [2, 2, 2]
+ num_filters: [64, 128, 256]
+ upsample_strides: [1, 2, 4]
+ num_upsample_filters: [128, 128, 128]
+ use_groupnorm: false
+ num_groups: 32
+ }
+ loss: {
+ classification_loss: {
+ weighted_sigmoid_focal: {
+ alpha: 0.25
+ gamma: 2.0
+ anchorwise_output: true
+ }
+ }
+ localization_loss: {
+ weighted_smooth_l1: {
+ sigma: 3.0
+ code_weight: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
+ }
+ }
+ classification_weight: 1.0
+ localization_weight: 2.0
+ }
+ # Outputs
+ use_sigmoid_score: true
+ encode_background_as_zeros: true
+ encode_rad_error_by_sin: true
+
+ use_direction_classifier: true
+ direction_loss_weight: 0.2
+ use_aux_classifier: false
+ # Loss
+ pos_class_weight: 1.0
+ neg_class_weight: 1.0
+
+ loss_norm_type: NormByNumPositives
+ # Postprocess
+ post_center_limit_range: [0, -39.68, -5, 69.12, 39.68, 5]
+ use_rotate_nms: false
+ use_multi_class_nms: false
+ nms_pre_max_size: 1000
+ nms_post_max_size: 300
+ nms_score_threshold: 0.3 #0.05
+ nms_iou_threshold: 0.1 #0.5
+
+ use_bev: false
+ num_point_features: 4
+ without_reflectivity: false
+ box_coder: {
+ ground_box3d_coder: {
+ linear_dim: false
+ encode_angle_vector: false
+ }
+ }
+ target_assigner: {
+ anchor_generators: {
+ anchor_generator_stride: {
+ sizes: [1.6, 3.9, 1.56] # wlh
+ strides: [0.32, 0.32, 0.0] # if generate only 1 z_center, z_stride will be ignored
+ offsets: [0.16, -39.52, -1.78] # origin_offset + strides / 2
+ rotations: [0, 1.57] # 0, pi/2
+ matched_threshold : 0.6
+ unmatched_threshold : 0.45
+ }
+ }
+
+ sample_positive_fraction : -1
+ sample_size : 512
+ region_similarity_calculator: {
+ nearest_iou_similarity: {
+ }
+ }
+ }
+ }
+}
+
+
+train_input_reader: {
+ record_file_path: "kitti_train.tfrecord"
+ class_names: ["Car"]
+ max_num_epochs : 160
+ batch_size: 1
+ prefetch_size : 25
+ max_number_of_voxels: 12000
+ shuffle_points: true
+ num_workers: 2
+ groundtruth_localization_noise_std: [0.25, 0.25, 0.25]
+ groundtruth_rotation_uniform_noise: [-0.15707963267, 0.15707963267]
+ global_rotation_uniform_noise: [-0.78539816, 0.78539816]
+ global_scaling_uniform_noise: [0.95, 1.05]
+ global_random_rotation_range_per_object: [0, 0]
+ anchor_area_threshold: 1
+ remove_points_after_sample: false
+ groundtruth_points_drop_percentage: 0.0
+ groundtruth_drop_max_keep_points: 15
+ database_sampler {
+ database_info_path: "kitti_dbinfos_train.pkl"
+ sample_groups {
+ name_to_max_num {
+ key: "Car"
+ value: 15
+ }
+ }
+ database_prep_steps {
+ filter_by_min_num_points {
+ min_num_point_pairs {
+ key: "Car"
+ value: 5
+ }
+ }
+ }
+ database_prep_steps {
+ filter_by_difficulty {
+ removed_difficulties: [-1]
+ }
+ }
+ global_random_rotation_range_per_object: [0, 0]
+ rate: 1.0
+ }
+
+ remove_unknown_examples: false
+ remove_environment: false
+ kitti_info_path: "kitti_infos_train.pkl"
+ kitti_root_path: ""
+}
+
+train_config: {
+
+ inter_op_parallelism_threads: 4
+ intra_op_parallelism_threads: 4
+ steps: 296960 # 1856 steps per epoch * 160 epochs
+ steps_per_eval: 9280 # 1856 steps per epoch * 5 epochs
+ save_checkpoints_secs : 1800 # half hour
+ save_summary_steps : 10
+ enable_mixed_precision: false
+ loss_scale_factor : 512.0
+ clear_metrics_every_epoch: false
+}
+
+eval_input_reader: {
+ record_file_path: "kitti_val.tfrecord"
+ class_names: ["Car"]
+ batch_size: 1
+ max_num_epochs : 160
+ prefetch_size : 25
+ max_number_of_voxels: 12000
+ shuffle_points: false
+ num_workers: 1
+ anchor_area_threshold: 1
+ remove_environment: false
+ kitti_info_path: "kitti_infos_val.pkl"
+ kitti_root_path: ""
+}
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/vnn_tapp_16.proto b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/vnn_tapp_16.proto
new file mode 100644
index 0000000000..23c5baae5b
--- /dev/null
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/configs/tanet/car/vnn_tapp_16.proto
@@ -0,0 +1,173 @@
+model: {
+ second: {
+ voxel_generator {
+ point_cloud_range : [0, -39.68, -3, 69.12, 39.68, 1]
+ voxel_size : [0.16, 0.16, 4]
+ max_number_of_points_per_voxel : 100
+ }
+ num_class: 1
+ voxel_feature_extractor: {
+ module_class_name: "PillarFeature_TANet"
+ num_filters: [64]
+ with_distance: false
+ }
+ middle_feature_extractor: {
+ module_class_name: "PointPillarsScatter"
+ }
+ rpn: {
+ module_class_name: "VRPN"
+ layer_nums: [3, 5, 5]
+ layer_strides: [2, 2, 2]
+ num_filters: [64, 128, 256]
+ upsample_strides: [1, 2, 4]
+ num_upsample_filters: [128, 128, 128]
+ use_groupnorm: false
+ num_groups: 32
+ }
+ loss: {
+ classification_loss: {
+ weighted_sigmoid_focal: {
+ alpha: 0.25
+ gamma: 2.0
+ anchorwise_output: true
+ }
+ }
+ localization_loss: {
+ weighted_smooth_l1: {
+ sigma: 3.0
+ code_weight: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
+ }
+ }
+ classification_weight: 1.0
+ localization_weight: 2.0
+ }
+ # Outputs
+ use_sigmoid_score: true
+ encode_background_as_zeros: true
+ encode_rad_error_by_sin: true
+
+ use_direction_classifier: true
+ direction_loss_weight: 0.2
+ use_aux_classifier: false
+ # Loss
+ pos_class_weight: 1.0
+ neg_class_weight: 1.0
+
+ loss_norm_type: NormByNumPositives
+ # Postprocess
+ post_center_limit_range: [0, -39.68, -5, 69.12, 39.68, 5]
+ use_rotate_nms: false
+ use_multi_class_nms: false
+ nms_pre_max_size: 1000
+ nms_post_max_size: 300
+ nms_score_threshold: 0.3 #0.05
+ nms_iou_threshold: 0.1 #0.5
+
+ use_bev: false
+ num_point_features: 4
+ without_reflectivity: false
+ box_coder: {
+ ground_box3d_coder: {
+ linear_dim: false
+ encode_angle_vector: false
+ }
+ }
+ target_assigner: {
+ anchor_generators: {
+ anchor_generator_stride: {
+ sizes: [1.6, 3.9, 1.56] # wlh
+ strides: [0.32, 0.32, 0.0] # if generate only 1 z_center, z_stride will be ignored
+ offsets: [0.16, -39.52, -1.78] # origin_offset + strides / 2
+ rotations: [0, 1.57] # 0, pi/2
+ matched_threshold : 0.6
+ unmatched_threshold : 0.45
+ }
+ }
+
+ sample_positive_fraction : -1
+ sample_size : 512
+ region_similarity_calculator: {
+ nearest_iou_similarity: {
+ }
+ }
+ }
+ }
+}
+
+
+train_input_reader: {
+ record_file_path: "kitti_train.tfrecord"
+ class_names: ["Car"]
+ max_num_epochs : 160
+ batch_size: 1
+ prefetch_size : 25
+ max_number_of_voxels: 12000
+ shuffle_points: true
+ num_workers: 2
+ groundtruth_localization_noise_std: [0.25, 0.25, 0.25]
+ groundtruth_rotation_uniform_noise: [-0.15707963267, 0.15707963267]
+ global_rotation_uniform_noise: [-0.78539816, 0.78539816]
+ global_scaling_uniform_noise: [0.95, 1.05]
+ global_random_rotation_range_per_object: [0, 0]
+ anchor_area_threshold: 1
+ remove_points_after_sample: false
+ groundtruth_points_drop_percentage: 0.0
+ groundtruth_drop_max_keep_points: 15
+ database_sampler {
+ database_info_path: "kitti_dbinfos_train.pkl"
+ sample_groups {
+ name_to_max_num {
+ key: "Car"
+ value: 15
+ }
+ }
+ database_prep_steps {
+ filter_by_min_num_points {
+ min_num_point_pairs {
+ key: "Car"
+ value: 5
+ }
+ }
+ }
+ database_prep_steps {
+ filter_by_difficulty {
+ removed_difficulties: [-1]
+ }
+ }
+ global_random_rotation_range_per_object: [0, 0]
+ rate: 1.0
+ }
+
+ remove_unknown_examples: false
+ remove_environment: false
+ kitti_info_path: "kitti_infos_train.pkl"
+ kitti_root_path: ""
+}
+
+train_config: {
+
+ inter_op_parallelism_threads: 4
+ intra_op_parallelism_threads: 4
+ steps: 296960 # 1856 steps per epoch * 160 epochs
+ steps_per_eval: 9280 # 1856 steps per epoch * 5 epochs
+ save_checkpoints_secs : 1800 # half hour
+ save_summary_steps : 10
+ enable_mixed_precision: false
+ loss_scale_factor : 512.0
+ clear_metrics_every_epoch: false
+}
+
+eval_input_reader: {
+ record_file_path: "kitti_val.tfrecord"
+ class_names: ["Car"]
+ batch_size: 1
+ max_num_epochs : 160
+ prefetch_size : 25
+ max_number_of_voxels: 12000
+ shuffle_points: false
+ num_workers: 1
+ anchor_area_threshold: 1
+ remove_environment: false
+ kitti_info_path: "kitti_infos_val.pkl"
+ kitti_root_path: ""
+}
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/data/kitti_common.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/data/kitti_common.py
index c074dcd736..9d15a92a9a 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/data/kitti_common.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/data/kitti_common.py
@@ -648,7 +648,7 @@ def get_pseudo_label_anno():
return annotations
-def get_start_result_anno():
+def get_start_result_anno(return_uncertainty):
annotations = {}
annotations.update({
"name": [],
@@ -661,6 +661,18 @@ def get_start_result_anno():
"rotation_y": [],
"score": [],
})
+
+ if return_uncertainty:
+ annotations.update({
+ "variance_truncated": [],
+ "variance_occluded": [],
+ "variance_alpha": [],
+ "variance_bbox": [],
+ "variance_dimensions": [],
+ "variance_location": [],
+ "variance_rotation_y": [],
+ "variance_score": [],
+ })
return annotations
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_coders.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_coders.py
index 0c3da0d1c2..d4d413394d 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_coders.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_coders.py
@@ -18,6 +18,10 @@ class GroundBox3dCoderTorch(GroundBox3dCoder):
return box_torch_ops.second_box_decode(
boxes, anchors, self.vec_encode, self.linear_dim
)
+ def decode_torch_uncertainty(self, boxes_var, boxes_mean, anchors):
+ return box_torch_ops.second_box_decode_uncertainty(
+ boxes_var, boxes_mean, anchors, self.vec_encode, self.linear_dim
+ )
class BevBoxCoderTorch(BevBoxCoder):
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_torch_ops.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_torch_ops.py
index 050ea5ae69..c36bea8317 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_torch_ops.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/core/box_torch_ops.py
@@ -102,6 +102,61 @@ def second_box_decode(box_encodings,
return torch.cat([xg, yg, zg, wg, lg, hg, rg], dim=-1)
+
+def second_box_decode_uncertainty(
+ box_var_encodings,
+ box_mean_encodings,
+ anchors,
+ encode_angle_to_vector=False,
+ smooth_dim=False):
+ """box decode for VoxelNet in lidar
+ Args:
+ boxes ([N, 7] Tensor): normal boxes: x, y, z, w, l, h, r
+ anchors ([N, 7] Tensor): anchors
+ """
+ xa, ya, za, wa, la, ha, ra = torch.split(anchors, 1, dim=-1)
+ if encode_angle_to_vector:
+ xt, yt, zt, wt, lt, ht, rtx, rty = torch.split(
+ box_var_encodings,
+ 1,
+ dim=-1
+ )
+ mxt, myt, mzt, mwt, mlt, mht, mrtx, mrty = torch.split(
+ box_mean_encodings,
+ 1,
+ dim=-1
+ )
+
+ else:
+ xt, yt, zt, wt, lt, ht, rt = torch.split(box_var_encodings, 1, dim=-1)
+ mxt, myt, mzt, mwt, mlt, mht, mrt = torch.split(box_mean_encodings, 1, dim=-1)
+
+ def exp_var(m, s):
+ return (torch.exp(s ** 2) - 1) * (torch.exp(2 * m + s ** 2))
+
+ za = za + ha / 2
+ diagonal = torch.sqrt(la**2 + wa**2)
+ xg = xt * diagonal ** 2
+ yg = yt * diagonal ** 2
+ zg = zt * ha ** 2
+ if smooth_dim:
+ lg = lt * la ** 2
+ wg = wt * wa ** 2
+ hg = ht * ha ** 2
+ else:
+ lg = exp_var(mlt, lt) * la ** 2
+ wg = exp_var(mwt, wt) * wa ** 2
+ hg = exp_var(mht, ht) * ha ** 2
+ if encode_angle_to_vector:
+ rgx = rtx
+ rgy = rty
+ rg = torch.atan2(rgy, rgx)
+ else:
+ rg = rt
+ return torch.cat([xg, yg, zg, wg, lg, hg, rg], dim=-1)
+
+
+
def bev_box_encode(boxes,
anchors,
encode_angle_to_vector=False,
@@ -372,6 +427,16 @@ def project_to_image(points_3d, proj_mat):
return point_2d_res
+def uncertainty_project_to_image(points_3d, proj_mat):
+ points_num = list(points_3d.shape)[:-1]
+ points_shape = np.concatenate([points_num, [1]], axis=0).tolist()
+ points_4 = torch.cat(
+ [points_3d, torch.zeros(*points_shape).type_as(points_3d)], dim=-1)
+ point_2d = points_4 @ (proj_mat.t() ** 2)
+ point_2d_res = point_2d[..., :2] / point_2d[..., 2:3]
+ return point_2d_res
+
+
def camera_to_lidar(points, r_rect, velo2cam):
num_points = points.shape[0]
points = torch.cat(
@@ -380,6 +445,17 @@ def camera_to_lidar(points, r_rect, velo2cam):
return lidar_points[..., :3]
+def uncertainty_camera_to_lidar(points, r_rect, velo2cam):
+ num_points = points.shape[0]
+ points = torch.cat(
+ [points, torch.ones(num_points, 1).type_as(points)], dim=-1)
+
+ M = torch.inverse((r_rect @ velo2cam).t())
+
+ lidar_points = points @ (M ** 2)
+ return lidar_points[..., :3]
+
+
def lidar_to_camera(points, r_rect, velo2cam):
num_points = points.shape[0]
points = torch.cat(
@@ -396,6 +472,14 @@ def box_camera_to_lidar(data, r_rect, velo2cam):
return torch.cat([xyz_lidar, w, l, h, r], dim=-1)
+def uncertainty_box_camera_to_lidar(data, r_rect, velo2cam):
+ xyz = data[..., 0:3]
+ l, h, w = data[..., 3:4], data[..., 4:5], data[..., 5:6]
+ r = data[..., 6:7]
+ xyz_lidar = uncertainty_camera_to_lidar(xyz, r_rect, velo2cam)
+ return torch.cat([xyz_lidar, w, l, h, r], dim=-1)
+
+
def box_lidar_to_camera(data, r_rect, velo2cam):
xyz_lidar = data[..., 0:3]
w, l, h = data[..., 3:4], data[..., 4:5], data[..., 5:6]
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/tanet.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/tanet.py
index b4d2bf2c8b..1131da76dd 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/tanet.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/tanet.py
@@ -930,7 +930,7 @@ class VPSA(nn.Module):
all_box_preds = []
all_cls_preds = []
all_dir_cls_preds = []
- all_refine_box_preds = []
+ all_refine_loc_preds = []
all_refine_cls_preds = []
all_refine_dir_cls_preds = []
@@ -994,8 +994,8 @@ class VPSA(nn.Module):
refine_cls_preds_local = self.refine_cls(PSA_output)
refine_loc_preds_local = self.refine_loc(PSA_output)
- all_refine_box_preds.append(refine_cls_preds_local)
- all_refine_cls_preds.append(refine_loc_preds_local)
+ all_refine_loc_preds.append(refine_loc_preds_local)
+ all_refine_cls_preds.append(refine_cls_preds_local)
if self._use_direction_classifier:
refine_dir_preds_local = self.refine_dir(PSA_output)
@@ -1021,8 +1021,8 @@ class VPSA(nn.Module):
torch.stack(all_cls_preds, dim=0), dim=0, unbiased=False
)
- refine_box_preds_var, refine_box_preds = torch.var_mean(
- torch.stack(all_refine_box_preds, dim=0), dim=0, unbiased=False
+ refine_loc_preds_var, refine_loc_preds = torch.var_mean(
+ torch.stack(all_refine_loc_preds, dim=0), dim=0, unbiased=False
)
refine_cls_preds_var, refine_cls_preds = torch.var_mean(
torch.stack(all_refine_cls_preds, dim=0), dim=0, unbiased=False
@@ -1033,9 +1033,9 @@ class VPSA(nn.Module):
cls_preds = cls_preds.permute(0, 2, 3, 1).contiguous()
box_preds_var = box_preds_var.permute(0, 2, 3, 1).contiguous()
cls_preds_var = cls_preds_var.permute(0, 2, 3, 1).contiguous()
- refine_box_preds = refine_box_preds.permute(0, 2, 3, 1).contiguous()
+ refine_loc_preds = refine_loc_preds.permute(0, 2, 3, 1).contiguous()
refine_cls_preds = refine_cls_preds.permute(0, 2, 3, 1).contiguous()
- refine_box_preds_var = refine_box_preds_var.permute(0, 2, 3, 1).contiguous()
+ refine_loc_preds_var = refine_loc_preds_var.permute(0, 2, 3, 1).contiguous()
refine_cls_preds_var = refine_cls_preds_var.permute(0, 2, 3, 1).contiguous()
ret_dict = {
@@ -1043,9 +1043,9 @@ class VPSA(nn.Module):
"cls_preds": cls_preds,
"box_preds_var": box_preds_var,
"cls_preds_var": cls_preds_var,
- "Refine_box_preds": refine_box_preds,
+ "Refine_loc_preds": refine_loc_preds,
"Refine_cls_preds": refine_cls_preds,
- "Refine_box_preds_var": refine_box_preds_var,
+ "Refine_loc_preds_var": refine_loc_preds_var,
"Refine_cls_preds_var": refine_cls_preds_var,
}
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/voxelnet.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/voxelnet.py
index bbee7d580b..cda9f04d66 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/voxelnet.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/pytorch/models/voxelnet.py
@@ -1100,7 +1100,7 @@ class VoxelNet(nn.Module):
def get_global_step(self):
return int(self.global_step.cpu().numpy()[0])
- def forward(self, example, refine_weight=2, samples=1):
+ def forward(self, example, refine_weight=2, samples=1, return_uncertainty=False):
"""module's forward should always accept dict and return loss.
"""
voxels = example["voxels"]
@@ -1303,18 +1303,21 @@ class VoxelNet(nn.Module):
"cared": cared,
}
else:
- if self.rpn_class_name == "PSA" or self.rpn_class_name == "RefineDet":
- coarse_output = self.predict_coarse(example, preds_dict, self.device)
- refine_output = self.predict_refine(example, preds_dict, self.device)
+ if self.rpn_class_name in ["PSA", "VPSA", "RefineDet"]:
+ coarse_output = self.predict_coarse(example, preds_dict, self.device, return_uncertainty=return_uncertainty)
+ refine_output = self.predict_refine(example, preds_dict, self.device, return_uncertainty=return_uncertainty)
return coarse_output, refine_output
else:
- return self.predict_coarse(example, preds_dict, self.device)
+ return self.predict_coarse(example, preds_dict, self.device, return_uncertainty=return_uncertainty)
def compute_predict(
self,
batch_box_preds,
batch_cls_preds,
batch_dir_preds,
+ batch_box_preds_var,
+ batch_cls_preds_var,
+ batch_dir_preds_var,
batch_rect,
batch_Trv2c,
batch_P2,
@@ -1322,12 +1325,16 @@ class VoxelNet(nn.Module):
batch_anchors_mask,
num_class_with_bg,
device,
+ return_uncertainty,
):
predictions_dicts = []
for (
box_preds,
cls_preds,
dir_preds,
+ box_preds_var,
+ cls_preds_var,
+ dir_preds_var,
rect,
Trv2c,
P2,
@@ -1337,6 +1344,9 @@ class VoxelNet(nn.Module):
batch_box_preds,
batch_cls_preds,
batch_dir_preds,
+ batch_box_preds_var,
+ batch_cls_preds_var,
+ batch_dir_preds_var,
batch_rect,
batch_Trv2c,
batch_P2,
@@ -1350,9 +1360,16 @@ class VoxelNet(nn.Module):
a_mask = a_mask.bool()
box_preds = box_preds[a_mask]
cls_preds = cls_preds[a_mask]
+
+ if return_uncertainty:
+ box_preds_var = box_preds_var[a_mask]
+ cls_preds_var = cls_preds_var[a_mask]
if self._use_direction_classifier:
if a_mask is not None:
dir_preds = dir_preds[a_mask]
+
+ if return_uncertainty:
+ dir_preds_var = dir_preds_var[a_mask]
dir_labels = torch.max(dir_preds, dim=-1)[1]
if self._encode_background_as_zeros:
# this don't support softmax
@@ -1398,27 +1415,45 @@ class VoxelNet(nn.Module):
)
selected_boxes, selected_labels, selected_scores = [], [], []
selected_dir_labels = []
+ selected_boxes_var = []
+ selected_cls_preds_var = []
+ selected_dir_preds_var = []
for i, selected in enumerate(selected_per_class):
if selected is not None:
num_dets = selected.shape[0]
selected_boxes.append(box_preds[selected])
selected_labels.append(
torch.full([num_dets], i, dtype=torch.int64))
+ selected_boxes.append(box_preds[selected])
if self._use_direction_classifier:
selected_dir_labels.append(dir_labels[selected])
selected_scores.append(total_scores[selected, i])
+
+ if return_uncertainty:
+ selected_boxes_var.append(box_preds_var[selected])
+ selected_dir_preds_var.append(dir_preds_var[selected])
+ selected_cls_preds_var.append(cls_preds_var[selected])
if len(selected_boxes) > 0:
selected_boxes = torch.cat(selected_boxes, dim=0)
selected_labels = torch.cat(selected_labels, dim=0)
selected_scores = torch.cat(selected_scores, dim=0)
+ if return_uncertainty:
+ selected_boxes_var = torch.cat(selected_boxes_var, dim=0)
+ selected_cls_preds_var = torch.cat(selected_cls_preds_var, dim=0)
if self._use_direction_classifier:
- selected_dir_labels = torch.cat(selected_dir_labels,
- dim=0)
+ selected_dir_labels = torch.cat(
+ selected_dir_labels,
+ dim=0
+ )
+ if return_uncertainty:
+ selected_dir_preds_var = torch.cat(selected_dir_preds_var, dim=0)
else:
selected_boxes = None
selected_labels = None
selected_scores = None
- selected_dir_labels = None
+ selected_boxes_var = None
+ selected_cls_preds_var = None
+ selected_dir_preds_var = None
else:
# get highest score per prediction, than apply nms
# to remove overlapped box.
@@ -1470,12 +1505,20 @@ class VoxelNet(nn.Module):
selected_dir_labels = dir_labels[selected]
selected_labels = top_labels[selected]
selected_scores = top_scores[selected]
+ if return_uncertainty:
+ selected_boxes_var = box_preds_var[selected]
+ selected_cls_preds_var = cls_preds_var[selected]
+ selected_dir_preds_var = dir_preds_var[selected]
# finally generate predictions.
if selected_boxes is not None:
box_preds = selected_boxes
scores = selected_scores
label_preds = selected_labels
+ if return_uncertainty:
+ box_preds_var = selected_boxes_var
+ cls_preds_var = selected_cls_preds_var
+ dir_preds_var = selected_dir_preds_var
if self._use_direction_classifier:
dir_labels = selected_dir_labels
opp_labels = dir_labels.byte() ^ (
@@ -1488,10 +1531,18 @@ class VoxelNet(nn.Module):
)
final_box_preds = box_preds
final_scores = scores
+ if return_uncertainty:
+ final_box_preds_var_camera = None
+ box_2d_preds_var = None
if is_calib:
final_box_preds_camera = box_torch_ops.box_lidar_to_camera(
final_box_preds, rect, Trv2c)
+
+ if return_uncertainty:
+ final_box_preds_var_camera = box_torch_ops.uncertainty_box_camera_to_lidar(
+ box_preds_var, rect, Trv2c
+ )
locs = final_box_preds_camera[:, :3]
dims = final_box_preds_camera[:, 3:6]
angles = final_box_preds_camera[:, 6]
@@ -1504,31 +1555,42 @@ class VoxelNet(nn.Module):
minxy = torch.min(box_corners_in_image, dim=1)[0]
maxxy = torch.max(box_corners_in_image, dim=1)[0]
box_2d_preds = torch.cat([minxy, maxxy], dim=1)
+
+ if return_uncertainty:
+ box_2d_preds_var = torch.zeros_like(box_2d_preds)
else:
box_2d_preds = None
final_box_preds_camera = None
# predictions
predictions_dict = {
"bbox": box_2d_preds,
+ "bbox_var": box_2d_preds_var if return_uncertainty else None,
"box3d_camera": final_box_preds_camera,
+ "box3d_camera_var": final_box_preds_var_camera if return_uncertainty else None,
"box3d_lidar": final_box_preds,
+ "box3d_lidar_var": box_preds_var if return_uncertainty else None,
"scores": final_scores,
+ "scores_var": cls_preds_var if return_uncertainty else None,
"label_preds": label_preds,
"image_idx": img_idx,
}
else:
predictions_dict = {
"bbox": None,
+ "bbox_var": None,
"box3d_camera": None,
+ "box3d_camera_var": None,
"box3d_lidar": None,
+ "box3d_lidar_var": None,
"scores": None,
+ "scores_var": None,
"label_preds": None,
"image_idx": img_idx,
}
predictions_dicts.append(predictions_dict)
return predictions_dicts
- def predict_coarse(self, example, preds_dict, device):
+ def predict_coarse(self, example, preds_dict, device, return_uncertainty):
t = time.time()
batch_size = example["anchors"].shape[0]
batch_anchors = example["anchors"].view(batch_size, -1, 7)
@@ -1556,38 +1618,79 @@ class VoxelNet(nn.Module):
t = time.time()
batch_box_preds = preds_dict["box_preds"]
batch_cls_preds = preds_dict["cls_preds"]
- batch_box_preds = batch_box_preds.view(batch_size, -1,
- self._box_coder.code_size)
+ batch_box_preds = batch_box_preds.view(
+ batch_size, -1,
+ self._box_coder.code_size
+ )
+
+ batch_box_preds_var = [None] * batch_size
+ batch_cls_preds_var = [None] * batch_size
+
+ if return_uncertainty:
+ batch_box_preds_var = preds_dict["box_preds_var"]
+ batch_cls_preds_var = preds_dict["cls_preds_var"]
+ batch_box_preds_var = batch_box_preds_var.view(
+ batch_size, -1,
+ self._box_coder.code_size
+ )
num_class_with_bg = self._num_class
if not self._encode_background_as_zeros:
num_class_with_bg = self._num_class + 1
- batch_cls_preds = batch_cls_preds.view(batch_size, -1,
- num_class_with_bg)
- batch_box_preds = self._box_coder.decode_torch(batch_box_preds,
- batch_anchors)
+ if return_uncertainty:
+ batch_cls_preds_var = batch_cls_preds_var.view(
+ batch_size, -1,
+ num_class_with_bg
+ )
+ batch_box_preds_var = self._box_coder.decode_torch_uncertainty(
+ batch_box_preds_var,
+ batch_box_preds,
+ batch_anchors
+ )
+
+ batch_cls_preds = batch_cls_preds.view(
+ batch_size, -1,
+ num_class_with_bg
+ )
+
+ batch_box_preds = self._box_coder.decode_torch(
+ batch_box_preds,
+ batch_anchors
+ )
+
if self._use_direction_classifier:
batch_dir_preds = preds_dict["dir_cls_preds"]
batch_dir_preds = batch_dir_preds.view(batch_size, -1, 2)
+
+ batch_dir_preds_var = [None] * batch_size
+
+ if return_uncertainty:
+ batch_dir_preds_var = preds_dict["dir_cls_preds_var"]
+ batch_dir_preds_var = batch_dir_preds_var.view(batch_size, -1, 2)
else:
batch_dir_preds = [None] * batch_size
+ batch_dir_preds_var = [None] * batch_size
predictions_dicts = self.compute_predict(
batch_box_preds,
batch_cls_preds,
batch_dir_preds,
+ batch_box_preds_var,
+ batch_cls_preds_var,
+ batch_dir_preds_var,
batch_rect,
batch_Trv2c,
batch_P2,
batch_imgidx,
batch_anchors_mask,
num_class_with_bg,
- device=device
+ device=device,
+ return_uncertainty=return_uncertainty,
)
self._total_postprocess_time += time.time() - t
return predictions_dicts
- def predict_refine(self, example, preds_dict, device):
+ def predict_refine(self, example, preds_dict, device, return_uncertainty):
t = time.time()
batch_size = example["anchors"].shape[0]
batch_anchors = example["anchors"].view(batch_size, -1, 7)
@@ -1623,39 +1726,89 @@ class VoxelNet(nn.Module):
refine_box_preds = preds_dict["Refine_loc_preds"]
refine_cls_preds = preds_dict["Refine_cls_preds"]
- coarse_box_preds = coarse_box_preds.view(batch_size, -1,
- self._box_coder.code_size)
+ refine_box_preds_var = [None] * batch_size
+ refine_cls_preds_var = [None] * batch_size
- refine_box_preds = refine_box_preds.view(batch_size, -1,
- self._box_coder.code_size)
+ if return_uncertainty:
+ refine_box_preds_var = preds_dict["Refine_loc_preds_var"]
+ refine_cls_preds_var = preds_dict["Refine_cls_preds_var"]
- de_coarse_boxes = self._box_coder.decode_torch(coarse_box_preds,
- batch_anchors)
- de_refine_boxes = self._box_coder.decode_torch(refine_box_preds,
- de_coarse_boxes)
+ coarse_box_preds = coarse_box_preds.view(
+ batch_size, -1,
+ self._box_coder.code_size
+ )
+
+ refine_box_preds = refine_box_preds.view(
+ batch_size, -1,
+ self._box_coder.code_size
+ )
+
+ if return_uncertainty:
+ refine_box_preds_var = refine_box_preds_var.view(
+ batch_size, -1,
+ self._box_coder.code_size
+ )
+
+ de_coarse_boxes = self._box_coder.decode_torch(
+ coarse_box_preds,
+ batch_anchors
+ )
+ de_refine_boxes = self._box_coder.decode_torch(
+ refine_box_preds,
+ de_coarse_boxes
+ )
+
+ if return_uncertainty:
+ de_refine_boxes_var = self._box_coder.decode_torch_uncertainty(
+ refine_box_preds_var,
+ refine_box_preds,
+ de_coarse_boxes
+ )
batch_box_preds = de_refine_boxes
batch_cls_preds = refine_cls_preds
- batch_cls_preds = batch_cls_preds.view(batch_size, -1,
- num_class_with_bg)
+ batch_cls_preds = batch_cls_preds.view(
+ batch_size, -1,
+ num_class_with_bg)
+
+ batch_cls_preds_var = [None] * batch_size
+ batch_box_preds_var = [None] * batch_size
+
+ if return_uncertainty:
+ batch_cls_preds_var = refine_cls_preds_var
+ batch_box_preds_var = de_refine_boxes_var
+ batch_cls_preds_var = batch_cls_preds_var.view(
+ batch_size, -1,
+ num_class_with_bg)
if self._use_direction_classifier:
batch_dir_preds = preds_dict["Refine_dir_preds"]
batch_dir_preds = batch_dir_preds.view(batch_size, -1, 2)
+
+ batch_dir_preds_var = [None] * batch_size
+
+ if return_uncertainty:
+ batch_dir_preds_var = preds_dict["Refine_dir_preds_var"]
+ batch_dir_preds_var = batch_dir_preds_var.view(batch_size, -1, 2)
else:
batch_dir_preds = [None] * batch_size
+ batch_dir_preds_var = [None] * batch_size
predictions_dicts = self.compute_predict(
batch_box_preds,
batch_cls_preds,
batch_dir_preds,
+ batch_box_preds_var,
+ batch_cls_preds_var,
+ batch_dir_preds_var,
batch_rect,
batch_Trv2c,
batch_P2,
batch_imgidx,
batch_anchors_mask,
num_class_with_bg,
- device=device
+ device=device,
+ return_uncertainty=return_uncertainty,
)
self._total_postprocess_time += time.time() - t
return predictions_dicts
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/run.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/run.py
index fa5c98f5da..6a1403982f 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/run.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/second_detector/run.py
@@ -606,6 +606,7 @@ def compute_lidar_kitti_output(
center_limit_range,
class_names,
global_set,
+ return_uncertainty,
):
annos = []
for i, preds_dict in enumerate(predictions_dicts):
@@ -613,10 +614,18 @@ def compute_lidar_kitti_output(
scores = preds_dict["scores"].detach().cpu().numpy()
box_preds_lidar = preds_dict["box3d_lidar"].detach().cpu().numpy()
label_preds = preds_dict["label_preds"].detach().cpu().numpy()
- anno = kitti.get_start_result_anno()
+
+ if return_uncertainty:
+ scores_var = preds_dict["scores_var"].detach().cpu().numpy()
+ box_preds_lidar_var = preds_dict["box3d_lidar_var"].detach().cpu().numpy()
+ else:
+ scores_var = np.empty_like(scores)
+ box_preds_lidar_var = np.empty_like(box_preds_lidar)
+
+ anno = kitti.get_start_result_anno(return_uncertainty=return_uncertainty)
num_example = 0
- for box_lidar, score, label in zip(
- box_preds_lidar, scores, label_preds
+ for box_lidar, box_lidar_var, score, score_var, label in zip(
+ box_preds_lidar, box_preds_lidar_var, scores, scores_var, label_preds
):
if center_limit_range is not None:
limit_range = np.array(center_limit_range)
@@ -631,6 +640,18 @@ def compute_lidar_kitti_output(
anno["dimensions"].append(box_lidar[3:6])
anno["location"].append(box_lidar[:3])
anno["rotation_y"].append(box_lidar[6])
+ anno["score"].append(score)
+
+ if return_uncertainty:
+ anno["variance_truncated"].append(0.0)
+ anno["variance_occluded"].append(0)
+ anno["variance_alpha"].append(0)
+ anno["variance_bbox"].append(None)
+ anno["variance_dimensions"].append(box_lidar_var[3:6])
+ anno["variance_location"].append(box_lidar_var[:3])
+ anno["variance_rotation_y"].append(box_lidar_var[6])
+ anno["variance_score"].append(np.mean(score_var))
+
if global_set is not None:
for i in range(100000):
if score in global_set:
@@ -638,7 +659,6 @@ def compute_lidar_kitti_output(
else:
global_set.add(score)
break
- anno["score"].append(score)
num_example += 1
if num_example != 0:
diff --git a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/voxel_object_detection_3d_learner.py b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/voxel_object_detection_3d_learner.py
index 85c5a401d3..9bd003712c 100644
--- a/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/voxel_object_detection_3d_learner.py
+++ b/src/opendr/perception/object_detection_3d/voxel_object_detection_3d/voxel_object_detection_3d_learner.py
@@ -25,7 +25,7 @@ from opendr.engine.datasets import (
ExternalDataset,
MappedDatasetIterator,
)
-from opendr.engine.data import PointCloud
+from opendr.engine.data import PointCloud, PointCloudWithCalibration
from opendr.perception.object_detection_3d.voxel_object_detection_3d.second_detector.load import (
create_model as second_create_model,
load_from_checkpoint,
@@ -55,7 +55,7 @@ from opendr.perception.object_detection_3d.voxel_object_detection_3d.second_dete
from opendr.perception.object_detection_3d.voxel_object_detection_3d.second_detector.data.preprocess import (
merge_second_batch,
)
-from opendr.engine.target import BoundingBox3DList
+from opendr.engine.target import BoundingBox3DList, UncertaintyBoundingBox3DList
from opendr.engine.constants import OPENDR_SERVER_URL
from urllib.request import urlretrieve
from urllib.error import URLError
@@ -97,6 +97,7 @@ class VoxelObjectDetection3DLearner(Learner):
"decay_factor": 0.8,
"staircase": True,
},
+ return_uncertainty=False,
):
# Pass the shared parameters on super's constructor so they can get initialized as class attributes
super(VoxelObjectDetection3DLearner, self).__init__(
@@ -122,6 +123,7 @@ class VoxelObjectDetection3DLearner(Learner):
self.model_dir = None
self.eval_checkpoint_dir = None
self.infer_point_cloud_mapper = None
+ self.calib_infer_point_cloud_mapper = None
if tanet_config_path is not None:
set_tanet_config(tanet_config_path)
@@ -131,6 +133,7 @@ class VoxelObjectDetection3DLearner(Learner):
self.model.rpn_ort_session = None # ONNX runtime inference session
self.input_config_prepared = False
self.eval_config_prepared = False
+ self.return_uncertainty = return_uncertainty
def save(self, path, verbose=False):
"""
@@ -431,7 +434,7 @@ class VoxelObjectDetection3DLearner(Learner):
return result
- def infer(self, point_clouds):
+ def infer(self, point_clouds, samples=1):
if self.model is None:
raise ValueError("No model loaded or created")
@@ -452,33 +455,72 @@ class VoxelObjectDetection3DLearner(Learner):
self.infer_point_cloud_mapper = infer_point_cloud_mapper
self.model.eval()
+ if self.calib_infer_point_cloud_mapper is None:
+
+ def create_map_point_cloud_dataset_func():
+
+ prep_func = create_prep_func(
+ self.input_config,
+ self.model_config,
+ False,
+ self.voxel_generator,
+ self.target_assigner,
+ use_sampler=False,
+ )
+
+ def map(point_cloud_with_calibration):
+
+ point_cloud = point_cloud_with_calibration.data
+ calib = point_cloud_with_calibration.calib
+
+ example = _prep_v9(point_cloud, calib, prep_func)
+
+ if point_cloud_with_calibration.image_shape is not None:
+ example["image_shape"] = point_cloud_with_calibration.image_shape
+
+ return example
+
+ return map
+
+ self.calib_infer_point_cloud_mapper = create_map_point_cloud_dataset_func()
+ self.model.eval()
+
input_data = None
- if isinstance(point_clouds, PointCloud):
- input_data = merge_second_batch(
- [self.infer_point_cloud_mapper(point_clouds.data)]
- )
+ def map_single(point_cloud):
+ if isinstance(point_clouds, PointCloudWithCalibration):
+ return self.calib_infer_point_cloud_mapper(point_cloud)
+ elif isinstance(point_clouds, PointCloud):
+ return self.infer_point_cloud_mapper(point_cloud.data)
+ else:
+ raise ValueError("PointCloud or PointCloudWithCalibration expected")
+
+ if isinstance(point_clouds, (PointCloud, PointCloudWithCalibration)):
+ input_data = merge_second_batch([map_single(point_clouds)])
elif isinstance(point_clouds, list):
input_data = merge_second_batch(
- [self.infer_point_cloud_mapper(x.data) for x in point_clouds]
+ [map_single(x) for x in point_clouds]
)
else:
return ValueError(
- "point_clouds should be a PointCloud or a list of PointCloud"
+ "point_clouds should be a PointCloud or a list of PointClouds"
)
output = self.model(
- example_convert_to_torch(input_data, self.float_dtype, device=self.device,)
+ example_convert_to_torch(input_data, self.float_dtype, device=self.device), samples=samples, return_uncertainty=self.return_uncertainty,
)
- if self.model_config.rpn.module_class_name == "PSA" or self.model_config.rpn.module_class_name == "RefineDet":
+ if self.model_config.rpn.module_class_name in ["PSA", "VPSA", "RefineDet"]:
output = output[-1]
annotations = compute_lidar_kitti_output(
- output, self.center_limit_range, self.class_names, None
+ output, self.center_limit_range, self.class_names, None, return_uncertainty=self.return_uncertainty
)
- result = [BoundingBox3DList.from_kitti(anno) for anno in annotations]
+ if self.return_uncertainty:
+ result = [UncertaintyBoundingBox3DList.from_kitti(anno) for anno in annotations]
+ else:
+ result = [BoundingBox3DList.from_kitti(anno) for anno in annotations]
if isinstance(point_clouds, PointCloud):
return result[0]
--
GitLab