mmdetection(1):FCOS代码解析

原文我就不贴了&＃xff0c;说一下感受吧&＃xff01;

从检测方法出来&＃xff0c;自我感觉一直不协调&＃xff0c;现有的方法如fast系列一直比较复杂&＃xff0c;强大的神经网络应该是简单高效的&＃xff0c;one-stage从yolo出来后感觉好了很多&＃xff0c;但是在最后的map上做roi anchor一直感觉特别冗余&＃xff0c;还好corner出来了&＃xff0c;但其也存在的问题&＃xff0c;不同的pool结构&＃xff0c;还有总感觉这种方式怪怪的。然后FCOS出来了&＃xff0c;完全感受到了高效和简单&＃xff0c;在此膜拜一下大神&＃xff0c;感觉神经网络就应该这样&＃xff0c;以最简单的方式&＃xff0c;取得很好的效果。

FCOS: Fully Convolutional One-Stage Object Detection

简单介绍一下&＃xff0c;方法很简单&＃xff0c;好文。

看到上图了吗&＃xff0c;没错&＃xff0c;就是这么粗暴&＃xff0c;像素级的预测t,l,r,b,当然&＃xff0c;上面存在着一个问题&＃xff0c;就是一个点包含两个框&＃xff0c;怎么搞&＃xff0c;当然选小的了&＃xff0c;不不不&＃xff0c;FPN啊&＃xff1a;

稍微解释一下最后的框框&＃xff0c;x4是4层卷积&＃xff0c;class分类&＃xff0c;距离中心的loss&＃xff0c;回归&＃xff0c;

然后是中心度的定义&＃xff0c;二进制交叉熵做loss。

ok&＃xff0c;讲完了&＃xff0c; 简单吧&＃xff0c;粗暴吧&＃xff0c;开始实现吧&＃xff01;

更新&＃xff1a;
非常好用个点FCOS得配点代码不是&＃xff0c;近期研究mmdetection,感觉很好用&＃xff0c;特来吧源码解释一下

def loss(self,cls_scores,bbox_preds,centernesses,gt_bboxes,gt_labels,img_metas,cfg,gt_bboxes_ignore&＃61;None):labels, bbox_targets &＃61; self.fcos_target(all_level_points, gt_bboxes,gt_labels) #生成labelflatten_cls_scores &＃61; [cls_score.permute(0, 2, 3, 1).reshape(-1, self.cls_out_channels)for cls_score in cls_scores]flatten_bbox_preds &＃61; [bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)for bbox_pred in bbox_preds]flatten_centerness &＃61; [centerness.permute(0, 2, 3, 1).reshape(-1)for centerness in centernesses]flatten_cls_scores &＃61; torch.cat(flatten_cls_scores)flatten_bbox_preds &＃61; torch.cat(flatten_bbox_preds)flatten_centerness &＃61; torch.cat(flatten_centerness)flatten_labels &＃61; torch.cat(labels)flatten_bbox_targets &＃61; torch.cat(bbox_targets)loss_cls &＃61; self.loss_cls(flatten_cls_scores, flatten_labels,avg_factor&＃61;num_pos &＃43; num_imgs) # cls losspos_bbox_preds &＃61; flatten_bbox_preds[pos_inds]pos_bbox_targets &＃61; flatten_bbox_targets[pos_inds]pos_centerness &＃61; flatten_centerness[pos_inds]pos_centerness_targets &＃61; self.centerness_target(pos_bbox_targets)pos_points &＃61; flatten_points[pos_inds]pos_decoded_bbox_preds &＃61; distance2bbox(pos_points, pos_bbox_preds)pos_decoded_target_preds &＃61; distance2bbox(pos_points,pos_bbox_targets)# centerness weighted iou lossloss_bbox &＃61; self.loss_bbox(pos_decoded_bbox_preds,pos_decoded_target_preds,weight&＃61;pos_centerness_targets,avg_factor&＃61;pos_centerness_targets.sum()) #llox lossloss_centerness &＃61; self.loss_centerness(pos_centerness,pos_centerness_targets) #center loss 交叉熵lossreturn dict(loss_cls&＃61;loss_cls,loss_bbox&＃61;loss_bbox,loss_centerness&＃61;loss_centerness)

其中生成中心点啊方法&＃xff0c;和原文一样&＃xff1a;

def centerness_target(self, pos_bbox_targets):# only calculate pos centerness targets, otherwise there may be nanleft_right &＃61; pos_bbox_targets[:, [0, 2]]top_bottom &＃61; pos_bbox_targets[:, [1, 3]]centerness_targets &＃61; (left_right.min(dim&＃61;-1)[0] / left_right.max(dim&＃61;-1)[0]) * (top_bottom.min(dim&＃61;-1)[0] / top_bottom.max(dim&＃61;-1)[0])return torch.sqrt(centerness_targets)


具体loss参考mmdet
focalloss&＃43;IOUloss&＃43; CEloss

bbox_head&＃61;dict(type&＃61;&＃39;FCOSHead&＃39;,num_classes&＃61;81,in_channels&＃61;256,stacked_convs&＃61;4,feat_channels&＃61;256,strides&＃61;[8, 16, 32, 64, 128],loss_cls&＃61;dict(type&＃61;&＃39;FocalLoss&＃39;,use_sigmoid&＃61;True,gamma&＃61;2.0,alpha&＃61;0.25,loss_weight&＃61;1.0),loss_bbox&＃61;dict(type&＃61;&＃39;IoULoss&＃39;, loss_weight&＃61;1.0),loss_centerness&＃61;dict(type&＃61;&＃39;CrossEntropyLoss&＃39;, use_sigmoid&＃61;True, loss_weight&＃61;1.0)))