def predict(self, example, output, pc_range):
t = time.time()
dets = ddd_decode(output['hm'],
output['rot'],
output['dim'],
pc_range,
example["ground"],
reg=output['reg'],
K=self.K)
batch_size = example['rect'].shape[0]
self._total_inference_count += batch_size
batch_rect = example["rect"]
batch_Trv2c = example["Trv2c"]
batch_P2 = example["P2"]
batch_imgidx = example['image_idx']
predictions_dicts = []
for det, rect, Trv2c, P2, img_idx in zip(dets, batch_rect, batch_Trv2c,
batch_P2, batch_imgidx):
final_alpha = get_alpha(det[:, 6:14])
final_rot_y = final_alpha + torch.atan2(-det[:, 1], det[:, 0])
final_box_preds = torch.cat(
[det[:, :6], final_rot_y.unsqueeze(1)], dim=-1)
final_scores = det[:, -2]
final_labels = det[:, -1]
final_box_preds_camera = box_torch_ops.box_lidar_to_camera(
final_box_preds, rect, Trv2c)
locs = final_box_preds_camera[:, :3]
dims = final_box_preds_camera[:, 3:6]
# change angles
angles = final_box_preds_camera[:, 6]
camera_box_origin = [0.5, 1.0, 0.5]
box_corners = box_torch_ops.center_to_corner_box3d(
locs, dims, angles, camera_box_origin, axis=1)
box_corners_in_image = box_torch_ops.project_to_image(
box_corners, P2)
# box_corners_in_image: [N, 8, 2]
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)
# predictions
predictions_dict = {
"bbox": box_2d_preds,
"box3d_camera": final_box_preds_camera,
"box3d_lidar": final_box_preds,
"scores": final_scores,
"label_preds": final_labels,
"image_idx": img_idx,
}
predictions_dicts.append(predictions_dict)
self._total_postprocess_time += time.time() - t
return predictions_dicts
def predict(self, example, preds_dict):
t = time.time()
batch_size = example['anchors'].shape[0]
batch_anchors = example["anchors"].view(batch_size, -1, 7)
self._total_inference_count += batch_size
batch_rect = example["rect"]
batch_Trv2c = example["Trv2c"]
batch_P2 = example["P2"]
if "anchors_mask" not in example:
batch_anchors_mask = [None] * batch_size
else:
batch_anchors_mask = example["anchors_mask"].view(batch_size, -1)
batch_imgidx = example['image_idx']
self._total_forward_time += time.time() - t
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)
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 self._use_direction_classifier:
batch_dir_preds = preds_dict["dir_cls_preds"]
batch_dir_preds = batch_dir_preds.view(batch_size, -1, 2)
else:
batch_dir_preds = [None] * batch_size
predictions_dicts = []
for box_preds, cls_preds, dir_preds, rect, Trv2c, P2, img_idx, a_mask in zip(
batch_box_preds, batch_cls_preds, batch_dir_preds, batch_rect,
batch_Trv2c, batch_P2, batch_imgidx, batch_anchors_mask):
if a_mask is not None:
box_preds = box_preds[a_mask]
cls_preds = cls_preds[a_mask]
if self._use_direction_classifier:
if a_mask is not None:
dir_preds = dir_preds[a_mask]
# print(dir_preds.shape)
dir_labels = torch.max(dir_preds, dim=-1)[1]
if self._encode_background_as_zeros:
# this don't support softmax
assert self._use_sigmoid_score is True
total_scores = torch.sigmoid(cls_preds)
else:
# encode background as first element in one-hot vector
if self._use_sigmoid_score:
total_scores = torch.sigmoid(cls_preds)[..., 1:]
else:
total_scores = F.softmax(cls_preds, dim=-1)[..., 1:]
# Apply NMS in birdeye view
if self._use_rotate_nms:
nms_func = box_torch_ops.rotate_nms
else:
nms_func = box_torch_ops.nms
selected_boxes = None
selected_labels = None
selected_scores = None
selected_dir_labels = None
if self._multiclass_nms:
# curently only support class-agnostic boxes.
boxes_for_nms = box_preds[:, [0, 1, 3, 4, 6]]
if not self._use_rotate_nms:
box_preds_corners = box_torch_ops.center_to_corner_box2d(
boxes_for_nms[:, :2], boxes_for_nms[:, 2:4],
boxes_for_nms[:, 4])
boxes_for_nms = box_torch_ops.corner_to_standup_nd(
box_preds_corners)
boxes_for_mcnms = boxes_for_nms.unsqueeze(1)
selected_per_class = box_torch_ops.multiclass_nms(
nms_func=nms_func,
boxes=boxes_for_mcnms,
scores=total_scores,
num_class=self._num_class,
pre_max_size=self._nms_pre_max_size,
post_max_size=self._nms_post_max_size,
iou_threshold=self._nms_iou_threshold,
score_thresh=self._nms_score_threshold,
)
selected_boxes, selected_labels, selected_scores = [], [], []
selected_dir_labels = []
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))
if self._use_direction_classifier:
selected_dir_labels.append(dir_labels[selected])
selected_scores.append(total_scores[selected, i])
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 self._use_direction_classifier:
selected_dir_labels = torch.cat(selected_dir_labels,
dim=0)
else:
selected_boxes = None
selected_labels = None
selected_scores = None
selected_dir_labels = None
else:
# get highest score per prediction, than apply nms
# to remove overlapped box.
if num_class_with_bg == 1:
top_scores = total_scores.squeeze(-1)
top_labels = torch.zeros(total_scores.shape[0],
device=total_scores.device,
dtype=torch.long)
else:
top_scores, top_labels = torch.max(total_scores, dim=-1)
if self._nms_score_threshold > 0.0:
thresh = torch.tensor(
[self._nms_score_threshold],
device=total_scores.device).type_as(total_scores)
top_scores_keep = (top_scores >= thresh)
top_scores = top_scores.masked_select(top_scores_keep)
if top_scores.shape[0] != 0:
if self._nms_score_threshold > 0.0:
box_preds = box_preds[top_scores_keep]
if self._use_direction_classifier:
dir_labels = dir_labels[top_scores_keep]
top_labels = top_labels[top_scores_keep]
boxes_for_nms = box_preds[:, [0, 1, 3, 4, 6]]
if not self._use_rotate_nms:
box_preds_corners = box_torch_ops.center_to_corner_box2d(
boxes_for_nms[:, :2], boxes_for_nms[:, 2:4],
boxes_for_nms[:, 4])
boxes_for_nms = box_torch_ops.corner_to_standup_nd(
box_preds_corners)
# the nms in 3d detection just remove overlap boxes.
selected = nms_func(
boxes_for_nms,
top_scores,
pre_max_size=self._nms_pre_max_size,
post_max_size=self._nms_post_max_size,
iou_threshold=self._nms_iou_threshold,
)
else:
selected = None
if selected is not None:
selected_boxes = box_preds[selected]
if self._use_direction_classifier:
selected_dir_labels = dir_labels[selected]
selected_labels = top_labels[selected]
selected_scores = top_scores[selected]
# finally generate predictions.
if selected_boxes is not None:
box_preds = selected_boxes
scores = selected_scores
label_preds = selected_labels
if self._use_direction_classifier:
dir_labels = selected_dir_labels
opp_labels = (box_preds[..., -1] > 0) ^ dir_labels.byte()
box_preds[..., -1] += torch.where(
opp_labels,
torch.tensor(np.pi).type_as(box_preds),
torch.tensor(0.0).type_as(box_preds))
# box_preds[..., -1] += (
# ~(dir_labels.byte())).type_as(box_preds) * np.pi
final_box_preds = box_preds
final_scores = scores
final_labels = label_preds
final_box_preds_camera = box_torch_ops.box_lidar_to_camera(
final_box_preds, rect, Trv2c)
locs = final_box_preds_camera[:, :3]
dims = final_box_preds_camera[:, 3:6]
angles = final_box_preds_camera[:, 6]
camera_box_origin = [0.5, 1.0, 0.5]
box_corners = box_torch_ops.center_to_corner_box3d(
locs, dims, angles, camera_box_origin, axis=1)
box_corners_in_image = box_torch_ops.project_to_image(
box_corners, P2)
# box_corners_in_image: [N, 8, 2]
minxy = torch.min(box_corners_in_image, dim=1)[0]
maxxy = torch.max(box_corners_in_image, dim=1)[0]
# minx = torch.min(box_corners_in_image[..., 0], dim=1)[0]
# maxx = torch.max(box_corners_in_image[..., 0], dim=1)[0]
# miny = torch.min(box_corners_in_image[..., 1], dim=1)[0]
# maxy = torch.max(box_corners_in_image[..., 1], dim=1)[0]
# box_2d_preds = torch.stack([minx, miny, maxx, maxy], dim=1)
box_2d_preds = torch.cat([minxy, maxxy], dim=1)
# predictions
predictions_dict = {
"bbox": box_2d_preds,
"box3d_camera": final_box_preds_camera,
"box3d_lidar": final_box_preds,
"scores": final_scores,
"label_preds": label_preds,
"image_idx": img_idx,
}
else:
predictions_dict = {
"bbox": None,
"box3d_camera": None,
"box3d_lidar": None,
"scores": None,
"label_preds": None,
"image_idx": img_idx,
}
predictions_dicts.append(predictions_dict)
self._total_postprocess_time += time.time() - t
return predictions_dicts
def get_projected_idx(input_size,
calib,
img_shape,
z_sel,
rot_noise,
scal_noise,
grid_size=4.,
right=False):
'''Compute anchor boxes for each feature map.
Args:
input_size: (tensor) model input size of (w,h).
Returns:
boxes: (list) anchor boxes for each feature map. Each of size [#anchors,4],
where #anchors = fmw * fmh * #anchors_per_cell
'''
## for FPN50 ##
# fm_sizes = [(input_size/pow(2.,i+3)).ceil() for i in range(self.num_fms)]
# grid_size = [8., 16., 32., 64., 128.]
## for PIXOR ##
fm_size = input_size
fm_w, fm_h = int(fm_size[0] / grid_size), int(fm_size[1] / grid_size)
xy2 = meshgrid(fm_w, fm_h).to(torch.float64) + 0.5
xy = (xy2 * grid_size).view(fm_w, fm_h, 1, 2).expand(fm_w, fm_h, 1, 2)
xy = xy.to(torch.float32)
z = torch.Tensor([z_sel]).view(1, 1, 1, 1).expand(fm_w, fm_h, 1, 1)
z = z.to(torch.float32)
box = torch.cat([xy, z], 3)
anchor_boxes = box.view(-1, 3)
# Calculate Anchor Center
anchor_center = torch.zeros(anchor_boxes.shape[0], 3, dtype=torch.float64)
# anchor_center[:, 0] = 70.4 - (anchor_boxes[:, 0] / 10) ## x
anchor_center[:, 0] = anchor_boxes[:, 0] / 10
anchor_center[:, 1] = (anchor_boxes[:, 1] / 10) - 40. ##y
anchor_center[:, 2] = anchor_boxes[:, 2] / 10
# Convert to velodyne coordinates
# anchor_center[:, 1] = -1 * anchor_center[:, 0]
# Adjust center_z to center from bottom
anchor_center[:, 2] += (1.52) / 2
# Apply inverse augmentation
# import pdb; pdb.set_trace()
anchor_center_np = anchor_center.numpy()
anchor_center_np = box_np_ops.rotation_points_single_angle(
anchor_center_np, -rot_noise, axis=2)
anchor_center_np *= 1. / scal_noise
# anchor_center_np = box_np_ops.rotation_points_single_angle(anchor_center_np, 1/scal_noise, axis=2)
# import pdb; pdb.set_trace()
anchor_center = torch.tensor(anchor_center_np, dtype=torch.float64)
# # Get GT height
# mask = ((max_ious>0.5)[0::2, ...].nonzero()*2).squeeze()
# anchor_center[mask, 2] = -1 * boxes_[max_ids[mask], 2].to(torch.float64)
# anchor_center[mask, 2] += (boxes_[max_ids[mask], 5].to(torch.float64)) / 2
# anchor_center = anchor_center[0::2, ...]
# Project to image space
# pts_2d, pts_2d_norm = anchor_projector.point_to_image(anchor_center, data_dir)
r_rect = torch.tensor(calib['rect'],
dtype=torch.float32,
device=torch.device("cpu")).to(torch.float64)
if right:
P2 = torch.tensor(calib['P3'],
dtype=torch.float32,
device=torch.device("cpu")).to(torch.float64)
else:
P2 = torch.tensor(calib['P2'],
dtype=torch.float32,
device=torch.device("cpu")).to(torch.float64)
velo2cam = torch.tensor(calib['Trv2c'],
dtype=torch.float32,
device=torch.device("cpu")).to(torch.float64)
# anchor_center = anchor_center[:,[1,0,2]]
anchor_center2 = box_torch_ops.lidar_to_camera(anchor_center, r_rect,
velo2cam)
idxs = box_torch_ops.project_to_image(anchor_center2, P2)
# image_h = img_shape[2] ##
# image_w = img_shape[1]
# img_shape_torch = torch.tensor([2496, 768]).to(torch.float64).view(1,2)
img_shape_torch = torch.tensor([1248, 384]).to(torch.float64).view(1, 2)
idxs_norm = idxs / img_shape_torch
# import pdb; pdb.set_trace()
# idx = idxs_norm
# # Filtering idx
# mask = torch.mul(idx > 0, idx < 1).sum(dim=1) == 2
# mask = mask.view(-1,1)
# import pdb; pdb.set_trace()
return idxs, idxs_norm
def train_stage_2(self, example, preds_dict, top_predictions_left,
top_predictions_right):
t = time.time()
batch_size = example['anchors'].shape[0]
batch_anchors = example["anchors"].view(batch_size, -1, 7)
batch_anchors_reshape = batch_anchors.reshape(1, 200, 176,
14) ## 预先设定的锚框？
batch_rect = example["rect"] ##
batch_Trv2c = example["Trv2c"]
batch_P2 = example["P2"]
batch_P3 = example["P3"]
batch_image_shape = example["image_shape"]
if "anchors_mask" not in example:
batch_anchors_mask = [None] * batch_size
else:
batch_anchors_mask = example["anchors_mask"].view(batch_size, -1)
batch_imgidx = example['image_idx']
t = time.time()
batch_box_preds = preds_dict["box_preds"] ## 预测的3d box
batch_cls_preds = preds_dict["cls_preds"] ## 预测的class
batch_box_preds = batch_box_preds.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 self._use_direction_classifier:
batch_dir_preds = preds_dict["dir_cls_preds"]
batch_dir_preds = batch_dir_preds.view(batch_size, -1, 2)
else:
batch_dir_preds = [None] * batch_size
predictions_dicts = []
for box_preds, cls_preds, dir_preds, rect, Trv2c, P2, P3, img_idx, a_mask in zip(
batch_box_preds, batch_cls_preds, batch_dir_preds, batch_rect,
batch_Trv2c, batch_P2, batch_P3, batch_imgidx,
batch_anchors_mask):
if a_mask is not None:
box_preds = box_preds[a_mask]
cls_preds = cls_preds[a_mask]
box_preds = box_preds.float()
cls_preds = cls_preds.float()
rect = rect.float()
Trv2c = Trv2c.float()
P2 = P2.float()
P3 = P3.float()
if self._encode_background_as_zeros:
# this don't support softmax
assert self._use_sigmoid_score is True
total_scores = torch.sigmoid(cls_preds)
#total_scores = cls_preds # use this if you want to fuse raw log score
else:
# encode background as first element in one-hot vector
if self._use_sigmoid_score:
total_scores = torch.sigmoid(cls_preds)[..., 1:]
else:
total_scores = F.softmax(cls_preds, dim=-1)[..., 1:]
# finally generate predictions.
final_box_preds = box_preds
final_scores = total_scores
final_box_preds_camera = box_torch_ops.box_lidar_to_camera(
final_box_preds, rect, Trv2c) ## 将3d box转换到图像坐标系
locs = final_box_preds_camera[:, :3]
dims = final_box_preds_camera[:, 3:6]
angles = final_box_preds_camera[:, 6]
camera_box_origin = [0.5, 1.0, 0.5]
box_corners = box_torch_ops.center_to_corner_box3d(
locs, dims, angles, camera_box_origin, axis=1) ##
box_corners_in_image_left = box_torch_ops.project_to_image(
box_corners, P2) ## 将8个顶点投影到图像
box_corners_in_image_right = box_torch_ops.project_to_image(
box_corners, P3)
# box_corners_in_image: [N, 8, 2]
minxy = torch.min(box_corners_in_image_left, dim=1)[0]
maxxy = torch.max(box_corners_in_image_left, dim=1)[0]
img_height = batch_image_shape[0, 0]
img_width = batch_image_shape[0, 1]
minxy[:, 0] = torch.clamp(minxy[:, 0], min=0, max=img_width)
minxy[:, 1] = torch.clamp(minxy[:, 1], min=0, max=img_height)
maxxy[:, 0] = torch.clamp(maxxy[:, 0], min=0, max=img_width)
maxxy[:, 1] = torch.clamp(maxxy[:, 1], min=0, max=img_height)
box_2d_preds_left = torch.cat([minxy, maxxy], dim=1)
minxy = torch.min(box_corners_in_image_right, dim=1)[0]
maxxy = torch.max(box_corners_in_image_right, dim=1)[0]
minxy[:, 0] = torch.clamp(minxy[:, 0], min=0, max=img_width)
minxy[:, 1] = torch.clamp(minxy[:, 1], min=0, max=img_height)
maxxy[:, 0] = torch.clamp(maxxy[:, 0], min=0, max=img_width)
maxxy[:, 1] = torch.clamp(maxxy[:, 1], min=0, max=img_height)
box_2d_preds_right = torch.cat([minxy, maxxy], dim=1)
# predictions
predictions_dict = {
"bbox": box_2d_preds_left,
"box3d_camera": final_box_preds_camera,
"box3d_lidar": final_box_preds,
"scores": final_scores,
#"label_preds": label_preds,
"image_idx": img_idx,
}
predictions_dicts.append(predictions_dict)
dis_to_lidar = torch.norm(
box_preds[:, :2], p=2, dim=1, keepdim=True) / 82.0 ## 到雷达的距离
box_2d_detector_left = np.zeros((200, 4))
box_2d_detector_right = np.zeros((200, 4))
#
if (top_predictions_left.shape[0] > 20):
box_2d_detector_left = top_predictions_left[:20, :4]
else:
box_2d_detector_left = top_predictions_left[:, :4]
if (top_predictions_right.shape[0] > 20):
box_2d_detector_right = top_predictions_right[:20, :4]
else:
box_2d_detector_right = top_predictions_right[:, :4]
# box_2d_detector[0:top_predictions.shape[0],:]=top_predictions[:,:4] ## 200个2d box
# box_2d_detector = top_predictions[:,:4]
# import ipdb;ipdb.set_trace()
box_2d_scores_left = top_predictions_left[:, 4].reshape(-1, 1)
box_2d_scores_right = top_predictions_right[:, 4].reshape(-1, 1)
time_iou_build_start = time.time()
overlaps = np.zeros((900000, 6),
dtype=box_2d_preds_left.detach().cpu().numpy().
dtype) ## 9x1e5个可能组合
tensor_index1 = np.zeros(
(900000, 2),
dtype=box_2d_preds_left.detach().cpu().numpy().dtype)
overlaps[:, :] = -1
tensor_index1[:, :] = -1
#final_scores[final_scores<0.1] = 0
#box_2d_preds[(final_scores<0.1).reshape(-1),:] = 0
iou_test, tensor_index, max_num = se.build_stage2_training(
box_2d_preds_left.detach().cpu().numpy(),
box_2d_preds_right.detach().cpu().numpy(),
box_2d_detector_left, box_2d_detector_right, -1,
final_scores.detach().cpu().numpy(), box_2d_scores_left,
box_2d_scores_right,
dis_to_lidar.detach().cpu().numpy(), overlaps, tensor_index1)
time_iou_build_end = time.time()
iou_test_tensor = torch.FloatTensor(
iou_test) #iou_test_tensor shape: [160000,4]
tensor_index_tensor = torch.LongTensor(tensor_index)
iou_test_tensor = iou_test_tensor.permute(1, 0)
iou_test_tensor = iou_test_tensor.reshape(1, 6, 1, 900000)
tensor_index_tensor = tensor_index_tensor.reshape(-1, 2)
if max_num == 0:
non_empty_iou_test_tensor = torch.zeros(1, 6, 1, 2)
non_empty_iou_test_tensor[:, :, :, :] = -1
non_empty_tensor_index_tensor = torch.zeros(2, 2)
non_empty_tensor_index_tensor[:, :] = -1
else:
non_empty_iou_test_tensor = iou_test_tensor[:, :, :, :max_num]
non_empty_tensor_index_tensor = tensor_index_tensor[:
max_num, :]
##
return predictions_dicts, non_empty_iou_test_tensor, non_empty_tensor_index_tensor
def rpn_nms(box_preds, cls_preds, example, box_coder, nms_score_threshold, nms_pre_max_size,
nms_post_max_size, nms_iou_threshold, training, range_thresh=0):
anchors = example["anchors"]
batch_size = anchors.shape[0]
batch_anchors = anchors.view(batch_size, -1, 7)
batch_rect = example["calib"]["rect"]
batch_Trv2c = example["calib"]["Trv2c"]
batch_P2 = example["calib"]["P2"]
if training:
batch_labels = example["labels"]
batch_reg_targets = example["reg_targets"]
batch_dir_targets = get_direction_target(
batch_anchors,
batch_reg_targets,
dir_offset=0.0,
num_bins=2)
else:
batch_labels = [None] * batch_size
batch_reg_targets = [None] * batch_size
batch_dir_targets = [None] * batch_size
if "anchors_mask" not in example:
batch_anchors_mask = [None] * batch_size
else:
anchors_mask = example["anchors_mask"]
batch_anchors_mask = anchors_mask.view(batch_size, -1)
batch_box_props = box_preds.view(batch_size, -1, box_coder.code_size)
batch_box_props = box_coder.decode_torch(batch_box_props, batch_anchors)
batch_cls_props = cls_preds.view(batch_size, -1, 1)
batch_far_proposals_bev = []
batch_far_proposals_img = []
batch_near_proposals_bev = []
batch_near_proposals_img = []
batch_rcnn_labels = []
batch_rcnn_reg_target = []
batch_rcnn_dir_target = []
batch_rcnn_anchors = []
for box_props, cls_props, labels, reg_target, dir_targets, rect, Trv2c, P2, a_mask, anchors in zip(
batch_box_props, batch_cls_props, batch_labels, batch_reg_targets, batch_dir_targets,
batch_rect, batch_Trv2c, batch_P2, batch_anchors_mask, batch_anchors):
if a_mask is not None:
box_props = box_props[a_mask]
cls_props = cls_props[a_mask]
anchors = anchors[a_mask]
if training:
labels = labels[a_mask]
reg_target = reg_target[a_mask]
dir_targets = dir_targets[a_mask]
cls_scores = torch.sigmoid(cls_props)[..., 1:]
top_scores = cls_props.squeeze(-1)
nms_func = box_torch_ops.nms
if nms_score_threshold > 0.0:
thresh = torch.Tensor([nms_score_threshold],
device=cls_scores.cpu().device).type_as(cls_scores)
top_scores_keep = (top_scores >= thresh)
top_scores = top_scores.masked_select(top_scores_keep)
if top_scores.shape[0] != 0:
# score threshold
if nms_score_threshold > 0.0:
box_props = box_props[top_scores_keep]
anchors = anchors[top_scores_keep]
if training:
labels = labels[top_scores_keep]
reg_target = reg_target[top_scores_keep]
dir_targets = dir_targets[top_scores_keep]
# range
range_thresh = torch.Tensor([range_thresh],
device=box_props.cpu().device).type_as(box_props)
# todo: uncertain, which is range
far_boxes_idx = (box_props[:, 0] >= range_thresh)
far_box_props = box_props[far_boxes_idx]
far_top_socres = top_scores[far_boxes_idx]
far_anchors = anchors[far_boxes_idx]
if training:
far_labels = labels[far_boxes_idx]
far_reg_target = reg_target[far_boxes_idx]
far_dir_target = dir_targets[far_boxes_idx]
if far_box_props.shape[0] != 0:
far_boxes_for_nms = far_box_props[:, [0, 1, 3, 4, 6]]
far_box_props_corners = box_torch_ops.center_to_corner_box2d(
far_boxes_for_nms[:, :2], far_boxes_for_nms[:, 2:4],
far_boxes_for_nms[:, 4])
far_boxes_for_nms = box_torch_ops.corner_to_standup_nd(
far_box_props_corners)
far_selected = nms_func(
far_boxes_for_nms,
far_top_socres,
pre_max_size=nms_pre_max_size // 2,
post_max_size=nms_post_max_size // 2,
iou_threshold=nms_iou_threshold)
else:
far_selected = None
if range_thresh > 0:
near_boxes_idx = (box_props[:, 0] < range_thresh)
near_box_props = box_props[near_boxes_idx]
near_anchors = anchors[near_boxes_idx]
near_top_socres = top_scores[near_boxes_idx]
if training:
near_labels = labels[near_boxes_idx]
near_reg_target = reg_target[near_boxes_idx]
near_dir_target = dir_targets[near_boxes_idx]
if near_box_props.shape[0] != 0:
near_boxes_for_nms = near_box_props[:, [0, 1, 3, 4, 6]]
near_box_props_corners = box_torch_ops.center_to_corner_box2d(
near_boxes_for_nms[:, :2], near_boxes_for_nms[:, 2:4],
near_boxes_for_nms[:, 4])
near_boxes_for_nms = box_torch_ops.corner_to_standup_nd(
near_box_props_corners)
near_selected = nms_func(
near_boxes_for_nms,
near_top_socres,
pre_max_size=nms_pre_max_size,
post_max_size=nms_post_max_size,
iou_threshold=nms_iou_threshold)
else:
near_selected = None
else:
near_selected = None
else:
far_selected = None
near_selected = None
if far_selected is not None:
far_proposals_3d = far_box_props[far_selected]
num_far_selected = far_proposals_3d.shape[0]
far_proposals_3d_fix = torch.zeros((nms_post_max_size // 2, 7)).cuda()
far_anchors_fix = torch.zeros((nms_post_max_size // 2, 7)).cuda()
far_proposals_3d_fix[:num_far_selected, :] = far_proposals_3d
far_anchors_fix[:num_far_selected, :] = far_anchors[far_selected]
far_anchors_fix = far_anchors_fix.unsqueeze(0)
if training:
far_labels_fix = torch.zeros((nms_post_max_size // 2)).cuda()
far_reg_target_fix = torch.zeros((nms_post_max_size // 2, 7)).cuda()
far_dir_target_fix = torch.zeros((nms_post_max_size // 2, 2)).cuda()
far_labels_fix[:num_far_selected] = far_labels[far_selected]
far_reg_target_fix[:num_far_selected, :] = far_reg_target[far_selected]
far_dir_target_fix[:num_far_selected, :] = far_dir_target[far_selected]
far_labels_fix = far_labels_fix.unsqueeze(0)
far_reg_target_fix = far_reg_target_fix.unsqueeze(0)
far_dir_target_fix = far_dir_target_fix.unsqueeze(0)
far_proposals_bev_fix = far_proposals_3d_fix[:, [0, 1, 3, 4, 6]].unsqueeze(0)
far_proposals_cam_fix = box_torch_ops.box_lidar_to_camera(far_proposals_3d_fix, rect, Trv2c)
far_locs_cam = far_proposals_cam_fix[:, :3]
far_dims_cam = far_proposals_cam_fix[:, 3:6]
far_angles_cam = far_proposals_cam_fix[:, 6]
camera_box_origin = [0.5, 1.0, 0.5]
far_proposals_cam_corners = box_torch_ops.center_to_corner_box3d(
far_locs_cam, far_dims_cam, far_angles_cam, camera_box_origin, axis=1)
far_proposals_img_corners = box_torch_ops.project_to_image(
far_proposals_cam_corners, P2)
minxy = torch.min(far_proposals_img_corners, dim=1)[0]
maxxy = torch.max(far_proposals_img_corners, dim=1)[0]
far_proposals_img_fix = torch.cat([minxy, maxxy], dim=1).unsqueeze(0)
else:
far_proposals_bev_fix = torch.zeros((nms_post_max_size // 2, 5)).cuda().unsqueeze(0)
far_proposals_img_fix = torch.zeros((nms_post_max_size // 2, 4)).cuda().unsqueeze(0)
far_labels_fix = torch.zeros((nms_post_max_size // 2)).cuda().unsqueeze(0)
far_reg_target_fix = torch.zeros((nms_post_max_size // 2, 7)).cuda().unsqueeze(0)
far_dir_target_fix = torch.zeros((nms_post_max_size // 2, 2)).cuda().unsqueeze(0)
far_anchors_fix = torch.zeros((nms_post_max_size // 2, 7)).cuda().unsqueeze(0)
if near_selected is not None:
near_proposals_3d = near_box_props[near_selected]
num_near_selected = near_proposals_3d.shape[0]
near_proposals_3d_fix = torch.zeros((nms_post_max_size, 7)).cuda()
near_anchors_fix = torch.zeros((nms_post_max_size, 7)).cuda()
near_proposals_3d_fix[:num_near_selected, :] = near_proposals_3d
near_anchors_fix[:num_near_selected, :] = near_anchors[near_selected]
near_anchors_fix = near_anchors_fix.unsqueeze(0)
if training:
near_labels_fix = torch.zeros((nms_post_max_size,)).cuda()
near_reg_target_fix = torch.zeros((nms_post_max_size, 7)).cuda()
near_dir_target_fix = torch.zeros((nms_post_max_size, 2)).cuda()
near_labels_fix[:num_near_selected] = near_labels[near_selected]
near_reg_target_fix[:num_near_selected, :] = near_reg_target[near_selected]
near_dir_target_fix[:num_near_selected, :] = near_dir_target[near_selected]
near_labels_fix = near_labels_fix.unsqueeze(0)
near_reg_target_fix = near_reg_target_fix.unsqueeze(0)
near_dir_target_fix = near_dir_target_fix.unsqueeze(0)
near_proposals_bev_fix = near_proposals_3d_fix[:, [0, 1, 3, 4, 6]].unsqueeze(0)
near_proposals_cam_fix = box_torch_ops.box_lidar_to_camera(near_proposals_3d_fix, rect, Trv2c)
near_locs_cam = near_proposals_cam_fix[:, :3]
near_dims_cam = near_proposals_cam_fix[:, 3:6]
near_angles_cam = near_proposals_cam_fix[:, 6]
camera_box_origin = [0.5, 1.0, 0.5]
near_proposals_cam_corners = box_torch_ops.center_to_corner_box3d(
near_locs_cam, near_dims_cam, near_angles_cam, camera_box_origin, axis=1)
near_proposals_img_corners = box_torch_ops.project_to_image(
near_proposals_cam_corners, P2)
near_minxy = torch.min(near_proposals_img_corners, dim=1)[0]
near_maxxy = torch.max(near_proposals_img_corners, dim=1)[0]
near_proposals_img_fix = torch.cat([near_minxy, near_maxxy], dim=1).unsqueeze(0)
else:
near_proposals_bev_fix = torch.zeros((nms_post_max_size, 5)).cuda().unsqueeze(0)
near_proposals_img_fix = torch.zeros((nms_post_max_size, 4)).cuda().unsqueeze(0)
near_labels_fix = torch.zeros((nms_post_max_size)).cuda().unsqueeze(0)
near_reg_target_fix = torch.zeros((nms_post_max_size, 7)).cuda().unsqueeze(0)
near_dir_target_fix = torch.zeros((nms_post_max_size, 2)).cuda().unsqueeze(0)
near_anchors_fix = torch.zeros((nms_post_max_size, 7)).cuda().unsqueeze(0)
if training:
rcnn_labels_fix = torch.cat([near_labels_fix, far_labels_fix], dim=1)
rcnn_reg_target_fix = torch.cat([near_reg_target_fix, far_reg_target_fix], dim=1)
rcnn_dir_target_fix = torch.cat([near_dir_target_fix, far_dir_target_fix], dim=1)
else:
rcnn_labels_fix = None
rcnn_reg_target_fix = None
rcnn_dir_target_fix = None
if near_anchors_fix is not None:
rcnn_anchors_fix = torch.cat([near_anchors_fix, far_anchors_fix], dim=1)
batch_far_proposals_bev.append(far_proposals_bev_fix)
batch_far_proposals_img.append(far_proposals_img_fix)
batch_near_proposals_bev.append(near_proposals_bev_fix)
batch_near_proposals_img.append(near_proposals_img_fix)
batch_rcnn_labels.append(rcnn_labels_fix)
batch_rcnn_reg_target.append(rcnn_reg_target_fix)
batch_rcnn_dir_target.append(rcnn_dir_target_fix)
batch_rcnn_anchors.append(rcnn_anchors_fix)
batch_far_proposals_bev = torch.cat(batch_far_proposals_bev, dim=0)
batch_far_proposals_img = torch.cat(batch_far_proposals_img, dim=0)
if batch_near_proposals_bev[0] is not None:
batch_near_proposals_bev = torch.cat(batch_near_proposals_bev, dim=0)
batch_near_proposals_img = torch.cat(batch_near_proposals_img, dim=0)
if training:
batch_rcnn_labels = torch.cat(batch_rcnn_labels, dim=0)
batch_rcnn_reg_target = torch.cat(batch_rcnn_reg_target, dim=0)
batch_rcnn_dir_target = torch.cat(batch_rcnn_dir_target, dim=0)
batch_rcnn_anchors = torch.cat(batch_rcnn_anchors, dim=0)
rcnn_examples = {
"far_props_bev": batch_far_proposals_bev,
"far_props_img": batch_far_proposals_img,
"near_props_bev": batch_near_proposals_bev,
"near_props_img": batch_near_proposals_img,
"rcnn_labels": batch_rcnn_labels,
"rcnn_reg_targets": batch_rcnn_reg_target,
"rcnn_dir_targets": batch_rcnn_dir_target,
"rcnn_anchors": batch_rcnn_anchors
}
return rcnn_examples
