def get_rescore_bboxes(self, guided_anchors, cls_scores, anchor_labels,
img_metas, cfg):
det_bboxes = list()
det_scores = list()
det_labels = list()
for i in range(len(img_metas)):
bbox_pred = guided_anchors[i]
scores = cls_scores[i]
labels = anchor_labels[i]
if scores.numel() == 0:
det_bboxes.append(None)
det_scores.append(None)
det_labels.append(None)
continue
bbox_pred = bbox_pred.view(-1, 7)
scores = torch.sigmoid(scores).view(-1)
select = scores > cfg.score_thr
bbox_pred = bbox_pred[select, :]
scores = scores[select]
labels = labels[select]
if scores.numel() == 0:
det_bboxes.append(None)
det_scores.append(None)
det_labels.append(None)
continue
boxes_for_nms = boxes3d_to_bev_torch(bbox_pred)
keep = rotate_nms_torch(boxes_for_nms,
scores,
iou_threshold=cfg.nms.iou_thr)
bbox_pred = bbox_pred[keep, :]
scores = scores[keep]
labels = labels[keep]
det_bboxes.append(bbox_pred.detach().cpu().numpy())
det_scores.append(scores.detach().cpu().numpy())
det_labels.append(labels.detach().cpu().numpy())
return det_bboxes, det_scores, det_labels
def get_det_bboxes(self, rois, cls_scores, bbox_preds, img_metas, cfg):
det_bboxes = list()
det_scores = list()
for i in range(len(img_metas)):
inds = torch.nonzero(rois[:, 0] == i).squeeze()
roi = rois[inds, 1:]
scores = cls_scores[inds, :]
if bbox_preds is not None:
bbox_pred = bbox_preds[inds, :]
else:
bbox_pred = roi
scores = torch.sigmoid(scores).view(-1)
select = scores > cfg.score_thr
bbox_pred = bbox_pred[select, :]
roi = roi[select, :]
scores = scores[select]
if scores.numel() == 0:
det_bboxes.append(bbox_pred)
det_scores.append(scores)
continue
if bbox_preds is not None:
bbox_pred = delta2rbbox3d(roi, bbox_pred, self.target_means,
self.target_stds)
boxes_for_nms = boxes3d_to_bev_torch(bbox_pred)
keep = rotate_nms_torch(boxes_for_nms,
scores,
iou_threshold=cfg.nms.iou_thr)
bbox_pred = bbox_pred[keep, :]
scores = scores[keep]
det_bboxes.append(bbox_pred.detach().cpu().numpy())
det_scores.append(scores.detach().cpu().numpy())
return det_bboxes, det_scores
def _get_proposals_single(self, rpn_cls_score, rpn_bbox_pred, rpn_dir_pred,
anchors, anchor_mask, img_meta, cfg):
if self.use_sigmoid_cls:
rpn_cls_score = rpn_cls_score.permute(1, 2,
0).contiguous().view(-1)
scores = rpn_cls_score.sigmoid()
else:
rpn_cls_score = rpn_cls_score.permute(1, 2,
0).contiguous().view(-1, 2)
scores = F.softmax(rpn_cls_score, dim=1)[:, 1]
rpn_bbox_pred = rpn_bbox_pred.permute(1, 2, 0).contiguous().view(-1, 7)
rpn_dir_label = torch.max(rpn_dir_pred.view(-1, 2), dim=-1)[1]
scores = scores[anchor_mask]
rpn_bbox_pred = rpn_bbox_pred[anchor_mask]
dir_label = rpn_dir_label[anchor_mask]
anchors = anchors[anchor_mask]
#####################################
# borrow from bbox_head, for expriments
# select = scores > .3
#
# bbox_pred = rpn_bbox_pred[select, :]
# anchors = anchors[select, :]
# scores = scores[select]
#
# if scores.numel() == 0:
# return bbox_pred, scores
#
# bboxes = delta2rbbox3d(anchors, bbox_pred, self.target_means,
# self.target_stds)
#
# keep = nms_gpu(
# boxes3d_to_bev_torch(bboxes), scores, .1)
# bboxes = bboxes.cpu().numpy()
# scores = scores.cpu().numpy()
# from mmdet.models.single_stage_heads.ssd_rotate_head import rotate_nms
# keep = rotate_nms(bboxes[:, [0, 1, 3, 4, 6]], scores, iou_threshold=.01)
# det_bboxes = bboxes[keep, :]
# det_scores = scores[keep]
#
# return (det_bboxes, det_scores)
#####################################
_, order = scores.sort(0, descending=True)
if cfg.nms_pre > 0:
order = order[:cfg.nms_pre]
rpn_bbox_pred = rpn_bbox_pred[order, :]
anchors = anchors[order, :]
scores = scores[order]
dir_label = dir_label[order]
proposals = delta2rbbox3d(anchors, rpn_bbox_pred, self.target_means,
self.target_stds)
keep = nms_gpu(boxes3d_to_bev_torch(proposals), scores, cfg.nms_thr)
proposals = proposals[keep, :]
scores = scores[keep]
dir_label = dir_label[keep]
proposals = proposals[:cfg.nms_post, :]
scores = scores[:cfg.nms_post]
dir_label = dir_label[:cfg.nms_post]
opp_labels = (proposals[..., -1] > 0) ^ dir_label.byte()
proposals[opp_labels, -1] += np.pi
return proposals