Closed dengyang1 closed 5 years ago
It can be proved that every point in the embedding corresponds to an 87*87 area in the search image. A 66 area in the embedding corresponds to a 127*127 area, the same size as the exemplar image. Two adjacent points in the embedding correspond to two 87*87 areas whose centers are at a distance of 8. The stride 8=2*2*2 originated from conv1, pool1 and pool2, whose stride are 2. A point in the response is obtained by a 6*6 area in the exemplar embedding correlating with a same size area in the search embedding, which means the similarity between exemplar image and an 127\127 area in the search image.
所以这个意思就是在response上的距离中心的差值乘以步长就是最终移动距离
You still need to divide it by upsample_factor and search_scale.
好的,非常感谢解答 @shuida
@shuida 打扰一下,假如知道在search image上的被追踪物体的长和宽,怎么才能算出对应到相应的原图片上的物体长和宽呢?
您好: 我想问个问题,就是在Inference阶段,在得到距离中心点的差值后,为啥要乘以网络步长: disp_instance_input = disp_instance_feat * self.model_config['embed_config']['stride'] 可能我不太明白这个算法
谢谢所做的工作。