Puppet97

> > > @yongqiangzhang1 > > > You can have a try. > > > [nms.zip](https://github.com/liuwei16/ALFNet/files/2426261/nms.zip) > > > > > > hi, @VideoObjectSearch , when i use the nms.zip,...

Retianet中的一些问题

2

感觉这里存在问题，Retinanet部分中，取了与gt重合度最大的iou的下标，然后将其对应的gt标签置为gt的标签，因为前面取了top_k个与anchor对应的gt，假设这里top_k=2,anchor数量为A，那么labels数量就是A*2，此时gt_assignment_for_gt中anchor的下标就和原来anchor下标无法对应了，因为anchor相当于取了top_k份，那么顺序就发生了改变 if config.allow_low_quality: labels[gt_assignment_for_gt] = gt_boxes_perimg[:, 4] low_quality_bbox_targets = bbox_transform_opr( anchors[gt_assignment_for_gt], gt_boxes_perimg[:, :4]) bbox_targets[gt_assignment_for_gt] = low_quality_bbox_targets

anchor生成的问题

1

为什么生成anchor时这里左上角和右上角为y1,x1,y2,x2? boxes = np.vstack((yv - anchor_size_y_2, xv - anchor_size_x_2, yv + anchor_size_y_2, xv + anchor_size_x_2))

请问为什么要计算真实框和预测框的iou，并将与其iou>阈值忽略掉，不太理解这段代码，谢谢

5

anch_ious = jaccard(gt_box, pred_boxes_for_ignore) anch_ious_max, _ = torch.max(anch_ious,dim=0) anch_ious_max = anch_ious_max.view(pred_boxes[i].size()[:3]) noobj_mask[i][anch_ious_max>self.ignore_threshold] = 0