def random_crop_frustum(bboxes,
rect,
Trv2c,
P2,
max_crop_height=1.0,
max_crop_width=0.9):
num_gt = bboxes.shape[0]
crop_minxy = np.random.uniform([1 - max_crop_width, 1 - max_crop_height],
[0.3, 0.3],
size=[num_gt, 2])
crop_maxxy = np.ones([num_gt, 2], dtype=bboxes.dtype)
crop_bboxes = np.concatenate([crop_minxy, crop_maxxy], axis=1)
left = np.random.choice([False, True], replace=False, p=[0.5, 0.5])
if left:
crop_bboxes[:, [0, 2]] -= crop_bboxes[:, 0:1]
# crop_relative_bboxes to real bboxes
crop_bboxes *= np.tile(bboxes[:, 2:] - bboxes[:, :2], [1, 2])
crop_bboxes += np.tile(bboxes[:, :2], [1, 2])
C, R, T = box_np_ops.projection_matrix_to_CRT_kitti(P2)
frustums = box_np_ops.get_frustum_v2(crop_bboxes, C)
frustums -= T
# frustums = np.linalg.inv(R) @ frustums.T
frustums = np.einsum("ij, akj->aki", np.linalg.inv(R), frustums)
frustums = box_np_ops.camera_to_lidar(frustums, rect, Trv2c)
return frustums
def sample_all(
self,
root_path,
gt_boxes,
gt_names,
num_point_features,
random_crop=False,
gt_group_ids=None,
calib=None,
road_planes=None,
):
sampled_num_dict = {}
sample_num_per_class = []
for class_name, max_sample_num in zip(self._sample_classes,
self._sample_max_nums):
sampled_num = int(max_sample_num -
np.sum([n == class_name for n in gt_names]))
sampled_num = np.round(self._rate * sampled_num).astype(np.int64)
sampled_num_dict[class_name] = sampled_num
sample_num_per_class.append(sampled_num)
sampled_groups = self._sample_classes
if self._use_group_sampling:
assert gt_group_ids is not None
sampled_groups = []
sample_num_per_class = []
for group_name, class_names in self._group_name_to_names:
sampled_nums_group = [sampled_num_dict[n] for n in class_names]
sampled_num = np.max(sampled_nums_group)
sample_num_per_class.append(sampled_num)
sampled_groups.append(group_name)
total_group_ids = gt_group_ids
sampled = []
sampled_gt_boxes = []
avoid_coll_boxes = gt_boxes
for class_name, sampled_num in zip(sampled_groups,
sample_num_per_class):
if sampled_num > 0:
if self._use_group_sampling:
sampled_cls = self.sample_group(class_name, sampled_num,
avoid_coll_boxes,
total_group_ids)
else:
sampled_cls = self.sample_class_v2(class_name, sampled_num,
avoid_coll_boxes)
sampled += sampled_cls
if len(sampled_cls) > 0:
if len(sampled_cls) == 1:
sampled_gt_box = sampled_cls[0]["box3d_lidar"][
np.newaxis, ...]
else:
sampled_gt_box = np.stack(
[s["box3d_lidar"] for s in sampled_cls], axis=0)
sampled_gt_boxes += [sampled_gt_box]
avoid_coll_boxes = np.concatenate(
[avoid_coll_boxes, sampled_gt_box], axis=0)
if self._use_group_sampling:
if len(sampled_cls) == 1:
sampled_group_ids = np.array(
sampled_cls[0]["group_id"])[np.newaxis, ...]
else:
sampled_group_ids = np.stack(
[s["group_id"] for s in sampled_cls], axis=0)
total_group_ids = np.concatenate(
[total_group_ids, sampled_group_ids], axis=0)
if len(sampled) > 0:
sampled_gt_boxes = np.concatenate(sampled_gt_boxes, axis=0)
if road_planes is not None:
# Only support KITTI
# image plane
assert False, "Not correct yet!"
a, b, c, d = road_planes
center = sampled_gt_boxes[:, :3]
center[:, 2] -= sampled_gt_boxes[:, 5] / 2
center_cam = box_np_ops.lidar_to_camera(
center, calib["rect"], calib["Trv2c"])
cur_height_cam = (-d - a * center_cam[:, 0] -
c * center_cam[:, 2]) / b
center_cam[:, 1] = cur_height_cam
lidar_tmp_point = box_np_ops.camera_to_lidar(
center_cam, calib["rect"], calib["Trv2c"])
cur_lidar_height = lidar_tmp_point[:, 2]
# botom to middle center
# kitti [0.5, 0.5, 0] center to [0.5, 0.5, 0.5]
sampled_gt_boxes[:,
2] = cur_lidar_height + sampled_gt_boxes[:,
5] / 2
# mv_height = sampled_gt_boxes[:, 2] - cur_lidar_height
# sampled_gt_boxes[:, 2] -= mv_height
num_sampled = len(sampled)
s_points_list = []
for info in sampled:
try:
# TODO fix point read error
s_points = np.fromfile(
str(pathlib.Path(root_path) / info["path"]),
dtype=np.float32).reshape(-1, num_point_features)
# if not add_rgb_to_points:
# s_points = s_points[:, :4]
if "rot_transform" in info:
rot = info["rot_transform"]
s_points[:, :
3] = box_np_ops.rotation_points_single_angle(
s_points[:, :4], rot, axis=2)
s_points[:, :3] += info["box3d_lidar"][:3]
s_points_list.append(s_points)
# print(pathlib.Path(info["path"]).stem)
except Exception:
print(str(pathlib.Path(root_path) / info["path"]))
continue
# gt_bboxes = np.stack([s["bbox"] for s in sampled], axis=0)
# if np.random.choice([False, True], replace=False, p=[0.3, 0.7]):
# do random crop.
if random_crop:
s_points_list_new = []
assert calib is not None
rect = calib["rect"]
Trv2c = calib["Trv2c"]
P2 = calib["P2"]
gt_bboxes = box_np_ops.box3d_to_bbox(sampled_gt_boxes, rect,
Trv2c, P2)
crop_frustums = prep.random_crop_frustum(
gt_bboxes, rect, Trv2c, P2)
for i in range(crop_frustums.shape[0]):
s_points = s_points_list[i]
mask = prep.mask_points_in_corners(
s_points, crop_frustums[i:i + 1]).reshape(-1)
num_remove = np.sum(mask)
if num_remove > 0 and (s_points.shape[0] -
num_remove) > 15:
s_points = s_points[np.logical_not(mask)]
s_points_list_new.append(s_points)
s_points_list = s_points_list_new
ret = {
"gt_names": np.array([s["name"] for s in sampled]),
"difficulty": np.array([s["difficulty"] for s in sampled]),
"gt_boxes": sampled_gt_boxes,
"points": np.concatenate(s_points_list, axis=0),
"gt_masks": np.ones((num_sampled, ), dtype=np.bool_),
}
if self._use_group_sampling:
ret["group_ids"] = np.array([s["group_id"] for s in sampled])
else:
ret["group_ids"] = np.arange(gt_boxes.shape[0],
gt_boxes.shape[0] + len(sampled))
else:
ret = None
return ret
def __call__(self, res, info):
res["mode"] = self.mode
if res["type"] in ["KittiDataset", "WaymoDataset", "LyftDataset"]:
points = res["lidar"]["points"]
elif res["type"] in ["NuScenesDataset"]:
points = res["lidar"]["combined"]
if self.mode == "train":
anno_dict = res["lidar"]["annotations"]
gt_dict = {
"gt_boxes": anno_dict["boxes"],
"gt_names": np.array(anno_dict["names"]).reshape(-1),
}
if "difficulty" not in anno_dict:
difficulty = np.zeros([anno_dict["boxes"].shape[0]],
dtype=np.int32)
gt_dict["difficulty"] = difficulty
else:
gt_dict["difficulty"] = anno_dict["difficulty"]
if "calib" in res:
calib = res["calib"]
else:
calib = None
if self.add_rgb_to_points:
assert calib is not None and "image" in res
image_path = res["image"]["image_path"]
image = (imgio.imread(str(pathlib.Path(root_path) /
image_path)).astype(np.float32) / 255)
points_rgb = box_np_ops.add_rgb_to_points(points, image,
calib["rect"],
calib["Trv2c"],
calib["P2"])
points = np.concatenate([points, points_rgb], axis=1)
num_point_features += 3
if self.reference_detections is not None:
assert calib is not None and "image" in res
C, R, T = box_np_ops.projection_matrix_to_CRT_kitti(P2)
frustums = box_np_ops.get_frustum_v2(reference_detections, C)
frustums -= T
frustums = np.einsum("ij, akj->aki", np.linalg.inv(R), frustums)
frustums = box_np_ops.camera_to_lidar(frustums, rect, Trv2c)
surfaces = box_np_ops.corner_to_surfaces_3d_jit(frustums)
masks = points_in_convex_polygon_3d_jit(points, surfaces)
points = points[masks.any(-1)]
if self.remove_outside_points:
assert calib is not None
image_shape = res["metadata"]["image_shape"]
points = box_np_ops.remove_outside_points(points, calib["rect"],
calib["Trv2c"],
calib["P2"], image_shape)
if self.remove_environment is True and self.mode == "train":
selected = keep_arrays_by_name(gt_names, target_assigner.classes)
_dict_select(gt_dict, selected)
masks = box_np_ops.points_in_rbbox(points, gt_dict["gt_boxes"])
points = points[masks.any(-1)]
if self.mode == "train":
selected = drop_arrays_by_name(gt_dict["gt_names"],
["DontCare", "ignore"])
_dict_select(gt_dict, selected)
if self.remove_unknown:
remove_mask = gt_dict["difficulty"] == -1
"""
gt_boxes_remove = gt_boxes[remove_mask]
gt_boxes_remove[:, 3:6] += 0.25
points = prep.remove_points_in_boxes(points, gt_boxes_remove)
"""
keep_mask = np.logical_not(remove_mask)
_dict_select(gt_dict, keep_mask)
gt_dict.pop("difficulty")
if self.min_points_in_gt > 0:
# points_count_rbbox takes 10ms with 10 sweeps nuscenes data
point_counts = box_np_ops.points_count_rbbox(
points, gt_dict["gt_boxes"])
mask = point_counts >= min_points_in_gt
_dict_select(gt_dict, mask)
gt_boxes_mask = np.array(
[n in self.class_names for n in gt_dict["gt_names"]],
dtype=np.bool_)
if self.db_sampler:
sampled_dict = self.db_sampler.sample_all(
res["metadata"]["image_prefix"],
gt_dict["gt_boxes"],
gt_dict["gt_names"],
res["metadata"]["num_point_features"],
self.random_crop,
gt_group_ids=None,
calib=calib,
road_planes=None # res["lidar"]["ground_plane"]
)
if sampled_dict is not None:
sampled_gt_names = sampled_dict["gt_names"]
sampled_gt_boxes = sampled_dict["gt_boxes"]
sampled_points = sampled_dict["points"]
sampled_gt_masks = sampled_dict["gt_masks"]
gt_dict["gt_names"] = np.concatenate(
[gt_dict["gt_names"], sampled_gt_names], axis=0)
gt_dict["gt_boxes"] = np.concatenate(
[gt_dict["gt_boxes"], sampled_gt_boxes])
gt_boxes_mask = np.concatenate(
[gt_boxes_mask, sampled_gt_masks], axis=0)
if self.remove_points_after_sample:
masks = box_np_ops.points_in_rbbox(
points, sampled_gt_boxes)
points = points[np.logical_not(masks.any(-1))]
points = np.concatenate([sampled_points, points], axis=0)
prep.noise_per_object_v3_(
gt_dict["gt_boxes"],
points,
gt_boxes_mask,
rotation_perturb=self.gt_rotation_noise,
center_noise_std=self.gt_loc_noise_std,
global_random_rot_range=self.global_random_rot_range,
group_ids=None,
num_try=100,
)
_dict_select(gt_dict, gt_boxes_mask)
gt_classes = np.array(
[self.class_names.index(n) + 1 for n in gt_dict["gt_names"]],
dtype=np.int32,
)
gt_dict["gt_classes"] = gt_classes
iskitti = res["type"] in ["KittiDataset"]
if self.kitti_double:
assert False, "No more KITTI"
gt_dict["gt_boxes"], points = prep.random_flip_both(
gt_dict["gt_boxes"], points, flip_coor=70.4 / 2)
elif self.flip_single or iskitti:
assert False, "nuscenes double flip is better"
gt_dict["gt_boxes"], points = prep.random_flip(
gt_dict["gt_boxes"], points)
else:
gt_dict["gt_boxes"], points = prep.random_flip_both(
gt_dict["gt_boxes"], points)
gt_dict["gt_boxes"], points = prep.global_rotation(
gt_dict["gt_boxes"],
points,
rotation=self.global_rotation_noise)
gt_dict["gt_boxes"], points = prep.global_scaling_v2(
gt_dict["gt_boxes"], points, *self.global_scaling_noise)
if self.shuffle_points:
# shuffle is a little slow.
np.random.shuffle(points)
if self.mode == "train" and self.random_select:
if self.npoints < points.shape[0]:
pts_depth = points[:, 2]
pts_near_flag = pts_depth < 40.0
far_idxs_choice = np.where(pts_near_flag == 0)[0]
near_idxs = np.where(pts_near_flag == 1)[0]
near_idxs_choice = np.random.choice(near_idxs,
self.npoints -
len(far_idxs_choice),
replace=False)
choice = (np.concatenate(
(near_idxs_choice, far_idxs_choice), axis=0)
if len(far_idxs_choice) > 0 else near_idxs_choice)
np.random.shuffle(choice)
else:
choice = np.arange(0, len(points), dtype=np.int32)
if self.npoints > len(points):
extra_choice = np.random.choice(choice,
self.npoints - len(points),
replace=False)
choice = np.concatenate((choice, extra_choice), axis=0)
np.random.shuffle(choice)
points = points[choice]
if self.symmetry_intensity:
points[:, -1] -= 0.5 # translate intensity to [-0.5, 0.5]
# points[:, -1] *= 2
if self.normalize_intensity and res["type"] in ["NuScenesDataset"]:
# print(points[:20, 3])
assert 0, "Velocity Accuracy drops 3 percent with normalization.."
points[:, 3] /= 255
res["lidar"]["points"] = points
if self.mode == "train":
res["lidar"]["annotations"] = gt_dict
return res, info
def __call__(self, res, info, res_1, info_1, res_2, info_2):
res["mode"] = self.mode
res_1["mode"] = self.mode
res_2["mode"] = self.mode
if res["type"] in ["KittiDataset", "LyftDataset", "LvxDataset"]:
points = res["lidar"]["points"]
points_1 = res_1["lidar"]["points"]
points_2 = res_2["lidar"]["points"]
elif res["type"] == "NuScenesDataset":
points = res["lidar"]["combined"]
if self.mode == "train":
anno_dict = res["lidar"]["annotations"]
gt_dict = {
"gt_boxes": anno_dict["boxes"],
# 加入之前之后一帧的boxes
"gt_boxes_1": anno_dict["boxes_1"],
"gt_boxes_2": anno_dict["boxes_2"],
"gt_boxes_3": anno_dict["boxes_3"],
"gt_names": np.array(anno_dict["names"]).reshape(-1),
}
if "difficulty" not in anno_dict:
difficulty = np.zeros([anno_dict["boxes"].shape[0]],
dtype=np.int32)
gt_dict["difficulty"] = difficulty
else:
gt_dict["difficulty"] = anno_dict["difficulty"]
if "calib" in res:
calib = res["calib"]
else:
calib = None
if self.add_rgb_to_points:
assert calib is not None and "image" in res
image_path = res["image"]["image_path"]
image = (imgio.imread(str(pathlib.Path(root_path) /
image_path)).astype(np.float32) / 255)
points_rgb = box_np_ops.add_rgb_to_points(points, image,
calib["rect"],
calib["Trv2c"],
calib["P2"])
points = np.concatenate([points, points_rgb], axis=1)
num_point_features += 3
if self.reference_detections is not None:
assert calib is not None and "image" in res
C, R, T = box_np_ops.projection_matrix_to_CRT_kitti(P2)
frustums = box_np_ops.get_frustum_v2(reference_detections, C)
frustums -= T
frustums = np.einsum("ij, akj->aki", np.linalg.inv(R), frustums)
frustums = box_np_ops.camera_to_lidar(frustums, rect, Trv2c)
surfaces = box_np_ops.corner_to_surfaces_3d_jit(frustums)
masks = points_in_convex_polygon_3d_jit(points, surfaces)
points = points[masks.any(-1)]
if self.remove_outside_points:
assert calib is not None
image_shape = res["image"]["image_shape"]
points = box_np_ops.remove_outside_points(points, calib["rect"],
calib["Trv2c"],
calib["P2"], image_shape)
if self.remove_environment is True and self.mode == "train":
selected = kitti.keep_arrays_by_name(gt_names,
target_assigner.classes)
_dict_select(gt_dict, selected)
masks = box_np_ops.points_in_rbbox(points, gt_dict["gt_boxes"])
points = points[masks.any(-1)]
if self.mode == "train":
selected = kitti.drop_arrays_by_name(gt_dict["gt_names"],
["DontCare", "ignore"])
_dict_select(gt_dict, selected)
if self.remove_unknown:
remove_mask = gt_dict["difficulty"] == -1
"""
gt_boxes_remove = gt_boxes[remove_mask]
gt_boxes_remove[:, 3:6] += 0.25
points = prep.remove_points_in_boxes(points, gt_boxes_remove)
"""
keep_mask = np.logical_not(remove_mask)
_dict_select(gt_dict, keep_mask)
gt_dict.pop("difficulty")
# 保证每个box内点的数量
if self.min_points_in_gt > 0:
# points_count_rbbox takes 10ms with 10 sweeps nuscenes data
point_counts = box_np_ops.points_count_rbbox(
points, gt_dict["gt_boxes"])
mask = point_counts >= min_points_in_gt
_dict_select(gt_dict, mask)
gt_boxes_mask = np.array(
[n in self.class_names for n in gt_dict["gt_names"]],
dtype=np.bool_)
# 数据增广暂时去掉，有点改不过来
if self.db_sampler:
pass
assert NotImplementedError
sampled_dict = self.db_sampler.sample_all(
res["metadata"]["image_prefix"],
gt_dict["gt_boxes"],
gt_dict["gt_names"],
res["metadata"]["num_point_features"],
self.random_crop,
gt_group_ids=None,
calib=calib,
)
if sampled_dict is not None:
sampled_gt_names = sampled_dict["gt_names"]
sampled_gt_boxes = sampled_dict["gt_boxes"]
sampled_points = sampled_dict["points"]
sampled_gt_masks = sampled_dict["gt_masks"]
gt_dict["gt_names"] = np.concatenate(
[gt_dict["gt_names"], sampled_gt_names], axis=0)
gt_dict["gt_boxes"] = np.concatenate(
[gt_dict["gt_boxes"], sampled_gt_boxes])
gt_boxes_mask = np.concatenate(
[gt_boxes_mask, sampled_gt_masks], axis=0)
if self.remove_points_after_sample:
masks = box_np_ops.points_in_rbbox(
points, sampled_gt_boxes)
points = points[np.logical_not(masks.any(-1))]
points = np.concatenate([sampled_points, points], axis=0)
# 暂时不加扰动了
# TODO:需要改4帧，t-2，t-1,t的点云，以及t+1的box
prep.noise_per_object_v3_(
gt_dict["gt_boxes"],
gt_dict["gt_boxes_1"],
gt_dict["gt_boxes_2"],
gt_dict["gt_boxes_3"],
points,
points_1,
points_2,
gt_boxes_mask,
rotation_perturb=self.gt_rotation_noise,
center_noise_std=self.gt_loc_noise_std,
global_random_rot_range=self.global_random_rot_range,
group_ids=None,
num_try=100,
)
_dict_select(gt_dict, gt_boxes_mask)
# 将类别转换为index
gt_classes = np.array(
[self.class_names.index(n) + 1 for n in gt_dict["gt_names"]],
dtype=np.int32,
)
gt_dict["gt_classes"] = gt_classes
# 添加了三帧的全局变换
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"], gt_dict[
"gt_boxes_2"], points, points_1, points_2 = prep.random_flip(
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"],
gt_dict["gt_boxes_2"], points, points_1, points_2)
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"], gt_dict[
"gt_boxes_2"], points, points_1, points_2 = prep.global_rotation(
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"],
gt_dict["gt_boxes_2"],
points,
points_1,
points_2,
rotation=self.global_rotation_noise)
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"], gt_dict[
"gt_boxes_2"], points, points_1, points_2 = prep.global_scaling_v2(
gt_dict["gt_boxes"], gt_dict["gt_boxes_1"],
gt_dict["gt_boxes_2"], points, points_1, points_2, *self.global_scaling_noise)
if self.shuffle_points:
# shuffle is a little slow.
np.random.shuffle(points)
np.random.shuffle(points_1)
np.random.shuffle(points_2)
if self.mode == "train" and self.random_select:
# 进行点的随机采样，暂时不需要
if self.npoints < points.shape[0]:
pts_depth = points[:, 2]
pts_near_flag = pts_depth < 40.0
far_idxs_choice = np.where(pts_near_flag == 0)[0]
near_idxs = np.where(pts_near_flag == 1)[0]
near_idxs_choice = np.random.choice(near_idxs,
self.npoints -
len(far_idxs_choice),
replace=False)
choice = (np.concatenate(
(near_idxs_choice, far_idxs_choice), axis=0)
if len(far_idxs_choice) > 0 else near_idxs_choice)
np.random.shuffle(choice)
else:
choice = np.arange(0, len(points), dtype=np.int32)
if self.npoints > len(points):
extra_choice = np.random.choice(choice,
self.npoints - len(points),
replace=False)
choice = np.concatenate((choice, extra_choice), axis=0)
np.random.shuffle(choice)
points = points[choice]
if self.symmetry_intensity:
points[:, -1] -= 0.5 # translate intensity to [-0.5, 0.5]
# points[:, -1] *= 2
res["lidar"]["points"] = points
res_1["lidar"]["points"] = points_1
res_2["lidar"]["points"] = points_2
if self.mode == "train":
import copy
res["lidar"]["annotations"] = gt_dict
res_1["lidar"]["annotations"] = copy.deepcopy(gt_dict)
res_2["lidar"]["annotations"] = copy.deepcopy(gt_dict)
return res, info, res_1,info_1,res_2,info_2