def main(base_dir):
for scene in tqdm.trange(21):
out_fname = base_dir / (str(scene).zfill(4) + ".csv")
pcl_dir = base_dir / str(scene).zfill(4)
pcl_dir.mkdir(exist_ok=True, parents=True)
scenes = []
track_ids = []
img_ids = []
num_poses = []
num_other_tracks = []
fnames = []
gt_poses = get_gt_poses_from_kitti(kitti_path=cfg.KITTI_PATH, scene=scene)
label_data = get_gt_detection_data_from_kitti(
kitti_path=cfg.KITTI_PATH, scene=scene, poses=gt_poses
)
for track_id, track_data in tqdm.tqdm(
label_data.items(), position=1, total=len(label_data)
):
for img_id, row_data in track_data.items():
pcl = _generate_point_cloud(row_data)
pcl_fname = pcl_dir / (
str(track_id).zfill(3) + str(img_id).zfill(4) + ".npy"
)
np.save(pcl_fname, pcl)
scenes.append(scene)
track_ids.append(track_id)
img_ids.append(img_id)
num_poses.append(max(track_data.keys()) - min(track_data.keys()))
num_other_tracks.append(
len(label_data)
) # not correct, but not needed for fake data
fnames.append(pcl_fname)
df = pd.DataFrame(
dict(
track_id=track_ids,
pointcloud_fname=fnames,
num_poses=num_poses,
num_other_tracks=num_other_tracks,
img_id=img_ids,
scene=scenes,
)
)
df.to_csv(out_fname, index=False)
flag_class = threading.Event
process_class = threading.Thread
shared_data = queue_class()
returned_data = queue_class()
writer_data_2d = queue_class()
writer_data_3d = queue_class()
slam_data = queue_class()
stop_flag = flag_class()
next_step = flag_class()
img_shape = get_image_shape(kitti_path, scene)
image_stream = _get_image_stream(kitti_path,
scene,
stop_flag,
offset=args.offset)
stereo_cam, T02 = get_cameras_from_kitti(kitti_path, scene)
gt_poses = get_gt_poses_from_kitti(kitti_path, scene)
detection_stream = get_detection_stream(
obj_detections_path,
scene=scene,
offset=args.offset,
object_ids=[int(idx)
for idx in args.indeces] if args.indeces else None,
classes=args.classes,
)
slam_process = process_class(target=_fake_slam,
args=[slam_data, gt_poses, args.offset],
name="Fake SLAM")
write_2d_process = process_class(
target=_write_2d_detections,
kwargs={
def setup(self, stage: Optional[str] = None):
all_files = list(
filter(lambda f: f.suffix == ".csv", self._dataset_dir.iterdir()))
if not all_files:
raise ValueError(f"No `.csv` files found at `{self._dataset_dir}`")
first_file = True
for f in all_files:
df = pd.read_csv(f, index_col=False)
if first_file:
dataset = df
first_file = False
else:
dataset = dataset.append(df)
dataset.dropna(inplace=True)
dataset.reset_index(inplace=True, drop=True)
target_yaw = []
target_pos = []
all_gt_data = {}
idx_to_remove = []
for idx, row in enumerate(dataset.itertuples()):
scene = int(row.scene)
if scene not in all_gt_data: # load gt data dynamically
gt_poses = get_gt_poses_from_kitti(
kitti_path=config.KITTI_PATH, scene=scene)
label_data = get_gt_detection_data_from_kitti(
kitti_path=config.KITTI_PATH, scene=scene, poses=gt_poses)
all_gt_data[scene] = label_data
img_id = int(row.img_id)
if self._track_id_mapping:
track_id = self._track_id_mapping.get(row.track_id)
if track_id is None:
idx_to_remove.append(idx)
continue
else:
track_id = int(row.track_id)
gt_track_data = all_gt_data[scene].get(track_id)
if gt_track_data is None:
idx_to_remove.append(idx)
continue
gt_data = gt_track_data.get(img_id)
if gt_data is None:
idx_to_remove.append(idx)
continue # no corresponding GT detection
target_yaw.append(gt_data.rot_angle)
target_pos.append(gt_data.cam_pos)
dataset.drop(idx_to_remove, inplace=True)
dataset["target_yaw"] = target_yaw
dataset["target_pos"] = target_pos
size = len(dataset)
val_size = int(
size *
(self._train_val_test_ratio[1] / sum(self._train_val_test_ratio)))
test_size = int(
size *
(self._train_val_test_ratio[2] / sum(self._train_val_test_ratio)))
train_size = size - val_size - test_size
# shuffle dataframe first
dataset = dataset.sample(frac=1, random_state=42)
self._dataset["train"] = dataset.iloc[:train_size]
self._dataset["val"] = dataset.iloc[train_size:train_size + val_size]
self._dataset["test"] = dataset.iloc[-test_size:]
