def sample_class_v2(self, name, num, gt_boxes):
'''
This func aims to select fixed number of gt boxes from gt database with collision (bev iou) test performed.
'''
# sample num gt_boxes from gt_database
sampled = self._sampler_dict[name].sample(num)
sampled = copy.deepcopy(sampled)
num_sampled = len(sampled)
num_gt = gt_boxes.shape[0]
# get all boxes: gt_boxes + sp_boxes
sp_boxes = np.stack([i["box3d_lidar"] for i in sampled],
axis=0) # todo: need modification here
boxes = np.concatenate([gt_boxes, sp_boxes], axis=0).copy()
offset = [0.0, 0.0]
if self.gt_aug_with_context > 0.0:
offset = [self.gt_aug_with_context, self.gt_aug_with_context]
# get all boxes_bev: gt_boxes_bev + sampled_boxes_bev
sp_boxes_new = boxes[num_gt:]
gt_boxes_bv = box_np_ops.center_to_corner_box2d(
gt_boxes[:, 0:2], gt_boxes[:, 3:5], gt_boxes[:, -1])
sp_boxes_bv = box_np_ops.center_to_corner_box2d(
sp_boxes_new[:, 0:2], sp_boxes_new[:, 3:5] + offset,
sp_boxes_new[:, -1])
total_bv = np.concatenate([gt_boxes_bv, sp_boxes_bv], axis=0)
# collision test on bev (stricter than 3d)
coll_mat = prep.box_collision_test(total_bv,
total_bv) # todo: too slow here
diag = np.arange(total_bv.shape[0])
coll_mat[diag, diag] = False
# get valid samples
valid_samples = []
for i in range(
num_gt, num_gt + num_sampled
): # todo: the overall box num may not meet the requirement
if coll_mat[i].any():
# i-th sampled box is not considered into auged gt-boxes.
coll_mat[i] = False
coll_mat[:, i] = False
else:
# i-th sampled box is considered into auged gt-boxes.
valid_samples.append(sampled[i - num_gt])
return valid_samples
def sample_class_v2(self, name, num, gt_boxes):
sampled = self._sampler_dict[name].sample(num)
sampled = copy.deepcopy(sampled)
num_gt = gt_boxes.shape[0]
num_sampled = len(sampled)
gt_boxes_bv = box_np_ops.center_to_corner_box2d(
gt_boxes[:, 0:2], gt_boxes[:, 3:5], gt_boxes[:, -1])
sp_boxes = np.stack([i["box3d_lidar"] for i in sampled], axis=0)
valid_mask = np.zeros([gt_boxes.shape[0]], dtype=np.bool_)
valid_mask = np.concatenate(
[valid_mask,
np.ones([sp_boxes.shape[0]], dtype=np.bool_)], axis=0)
boxes = np.concatenate([gt_boxes, sp_boxes], axis=0).copy()
if self._enable_global_rot:
# place samples to any place in a circle.
prep.noise_per_object_v3_(boxes,
None,
valid_mask,
0,
0,
self._global_rot_range,
num_try=100)
sp_boxes_new = boxes[gt_boxes.shape[0]:]
sp_boxes_bv = box_np_ops.center_to_corner_box2d(
sp_boxes_new[:, 0:2], sp_boxes_new[:, 3:5], sp_boxes_new[:, -1])
total_bv = np.concatenate([gt_boxes_bv, sp_boxes_bv], axis=0)
# coll_mat = collision_test_allbox(total_bv)
coll_mat = prep.box_collision_test(total_bv, total_bv)
diag = np.arange(total_bv.shape[0])
coll_mat[diag, diag] = False
valid_samples = []
for i in range(num_gt, num_gt + num_sampled):
if coll_mat[i].any():
coll_mat[i] = False
coll_mat[:, i] = False
else:
if self._enable_global_rot:
sampled[i - num_gt]["box3d_lidar"][:2] = boxes[i, :2]
sampled[i - num_gt]["box3d_lidar"][-1] = boxes[i, -1]
sampled[i - num_gt]["rot_transform"] = (
boxes[i, -1] - sp_boxes[i - num_gt, -1])
valid_samples.append(sampled[i - num_gt])
return valid_samples
def rotate_nms_cc(dets, thresh):
scores = dets[:, 5]
order = scores.argsort()[::-1].astype(np.int32) # highest->lowest
dets_corners = box_np_ops.center_to_corner_box2d(
dets[:, :2], dets[:, 2:4], dets[:, 4]
)
dets_standup = box_np_ops.corner_to_standup_nd(dets_corners)
standup_iou = box_np_ops.iou_jit(dets_standup, dets_standup, eps=0.0)
# print(dets_corners.shape, order.shape, standup_iou.shape)
return rotate_non_max_suppression_cpu(dets_corners, order, standup_iou, thresh)
def filter_gt_box_outside_range(gt_boxes, limit_range):
"""remove gtbox outside training range.
this function should be applied after other prep functions
Args:
gt_boxes ([type]): [description]
limit_range ([type]): [description]
"""
gt_boxes_bv = box_np_ops.center_to_corner_box2d(gt_boxes[:, [0, 1]],
gt_boxes[:, [3, 3 + 1]],
gt_boxes[:, -1])
bounding_box = box_np_ops.minmax_to_corner_2d(
np.asarray(limit_range)[np.newaxis, ...])
ret = points_in_convex_polygon_jit(gt_boxes_bv.reshape(-1, 2),
bounding_box)
return np.any(ret.reshape(-1, 4), axis=1)
def rotate_weighted_nms_cc(box,
dets,
thresh,
iou_preds,
labels,
dirs,
anchors=None,
nms_cnt_thresh=2.6,
nms_sigma_dist_interval=(0, 20, 40, 60),
nms_sigma_square=(0.0009, 0.009, 0.1, 1),
suppressed_thresh=0.3,
):
scores = dets[:, 5]
dets_corners = box_np_ops.center_to_corner_box2d(
dets[:, :2], dets[:, 2:4], dets[:, 4]
)
dets_standup = box_np_ops.corner_to_standup_nd(dets_corners)
standup_iou = box_np_ops.iou_jit(dets_standup, dets_standup, eps=0.0)
if anchors is None:
centerness_c = 0
anchors = np.zeros((1,1))
else:
centerness_c = 1
result = IOU_weighted_rotate_non_max_suppression_cpu(box,
dets_corners,
standup_iou,
thresh,
scores,
iou_preds,
labels,
dirs,
anchors,
nms_cnt_thresh,
nms_sigma_dist_interval,
nms_sigma_square,
suppressed_thresh,
centerness_c)
return result
def rotate_weighted_nms_cc(
box,
dets,
thresh,
iou_preds,
labels,
dirs,
anchors=None,
):
scores = dets[:, 5]
order = scores.argsort()[::-1].astype(np.int32) # highest->lowest
dets_corners = box_np_ops.center_to_corner_box2d(dets[:, :2], dets[:, 2:4],
dets[:, 4])
dets_standup = box_np_ops.corner_to_standup_nd(dets_corners)
standup_iou = box_np_ops.iou_jit(dets_standup, dets_standup, eps=0.0)
result = IOU_weighted_rotate_non_max_suppression_cpu(
box, dets_corners, standup_iou, thresh, scores, iou_preds, labels,
dirs, anchors)
return result
def sample_group(self, name, num, gt_boxes, gt_group_ids):
sampled, group_num = self.sample(name, num)
sampled = copy.deepcopy(sampled)
# rewrite sampled group id to avoid duplicated with gt group ids
gid_map = {}
max_gt_gid = np.max(gt_group_ids)
sampled_gid = max_gt_gid + 1
for s in sampled:
gid = s["group_id"]
if gid in gid_map:
s["group_id"] = gid_map[gid]
else:
gid_map[gid] = sampled_gid
s["group_id"] = sampled_gid
sampled_gid += 1
num_gt = gt_boxes.shape[0]
gt_boxes_bv = box_np_ops.center_to_corner_box2d(
gt_boxes[:, 0:2], gt_boxes[:, 3:5], gt_boxes[:, -1])
sp_boxes = np.stack([i["box3d_lidar"] for i in sampled], axis=0)
sp_group_ids = np.stack([i["group_id"] for i in sampled], axis=0)
valid_mask = np.zeros([gt_boxes.shape[0]], dtype=np.bool_)
valid_mask = np.concatenate(
[valid_mask,
np.ones([sp_boxes.shape[0]], dtype=np.bool_)], axis=0)
boxes = np.concatenate([gt_boxes, sp_boxes], axis=0).copy()
group_ids = np.concatenate([gt_group_ids, sp_group_ids], axis=0)
if self._enable_global_rot:
# place samples to any place in a circle.
prep.noise_per_object_v3_(
boxes,
None,
valid_mask,
0,
0,
self._global_rot_range,
group_ids=group_ids,
num_try=100,
)
sp_boxes_new = boxes[gt_boxes.shape[0]:]
sp_boxes_bv = box_np_ops.center_to_corner_box2d(
sp_boxes_new[:, 0:2], sp_boxes_new[:, 3:5], sp_boxes_new[:, -1])
total_bv = np.concatenate([gt_boxes_bv, sp_boxes_bv], axis=0)
# coll_mat = collision_test_allbox(total_bv)
coll_mat = prep.box_collision_test(total_bv, total_bv)
diag = np.arange(total_bv.shape[0])
coll_mat[diag, diag] = False
valid_samples = []
idx = num_gt
for num in group_num:
if coll_mat[idx:idx + num].any():
coll_mat[idx:idx + num] = False
coll_mat[:, idx:idx + num] = False
else:
for i in range(num):
if self._enable_global_rot:
sampled[idx - num_gt +
i]["box3d_lidar"][:2] = boxes[idx + i, :2]
sampled[idx - num_gt +
i]["box3d_lidar"][-1] = boxes[idx + i, -1]
sampled[idx - num_gt + i]["rot_transform"] = (
boxes[idx + i, -1] -
sp_boxes[idx + i - num_gt, -1])
valid_samples.append(sampled[idx - num_gt + i])
idx += num
return valid_samples
def sample_class_v2(self, name, num, gt_boxes):
# sample num gt_boxes from gt_database
sampled = self._sampler_dict[name].sample(num)
sampled = copy.deepcopy(sampled)
num_sampled = len(sampled)
# get all boxes: gt_boxes + sp_boxes
num_gt = gt_boxes.shape[0]
gt_boxes_bv = box_np_ops.center_to_corner_box2d(
gt_boxes[:, 0:2], gt_boxes[:, 3:5], gt_boxes[:, -1])
sp_boxes = np.stack([i["box3d_lidar"] for i in sampled],
axis=0) # todo: need modification here
boxes = np.concatenate([gt_boxes, sp_boxes], axis=0).copy()
# unused: get mask
valid_mask = np.zeros([gt_boxes.shape[0]], dtype=np.bool_)
valid_mask = np.concatenate(
[valid_mask,
np.ones([sp_boxes.shape[0]], dtype=np.bool_)], axis=0)
if self._enable_global_rot: # False
prep.noise_per_object_v3_(boxes,
None,
valid_mask,
0,
0,
self._global_rot_range,
num_try=100)
# get all boxes_bev: gt_boxes_bev + sampled_boxes_bev
sp_boxes_new = boxes[gt_boxes.shape[0]:]
sp_boxes_bv = box_np_ops.center_to_corner_box2d(
sp_boxes_new[:, 0:2], sp_boxes_new[:, 3:5], sp_boxes_new[:, -1])
total_bv = np.concatenate([gt_boxes_bv, sp_boxes_bv], axis=0)
# collision test on bev (stricter than 3d)
coll_mat = prep.box_collision_test(total_bv,
total_bv) # todo: too slow here
diag = np.arange(total_bv.shape[0])
coll_mat[diag, diag] = False
# get valid samples
valid_samples = []
for i in range(
num_gt, num_gt + num_sampled
): # todo: without multiple try times, sometimes, the overall box num may not meet the requirement
if coll_mat[i].any():
# i-th sampled box is not considered into gt-boxes.
coll_mat[i] = False
coll_mat[:, i] = False
else:
if self._enable_global_rot: # False
sampled[i - num_gt]["box3d_lidar"][:2] = boxes[
i, :2] # boxes has got noise
sampled[i - num_gt]["box3d_lidar"][-1] = boxes[i, -1]
sampled[i - num_gt]["rot_transform"] = (
boxes[i, -1] - sp_boxes[i - num_gt, -1])
# i-th sampled box is considered into gt-boxes.
valid_samples.append(sampled[i - num_gt])
return valid_samples
