taohan10200 / IIM

PyTorch implementations of the paper: "Learning Independent Instance Maps for Crowd Localization"
MIT License
163 stars 39 forks source link

Great work but Threshold Encoder module is useless.Details look at the picture #24

Open lab-gpu opened 2 years ago

lab-gpu commented 2 years ago

Thanks for the author's work.Great work!!!!! 1:Working so hard to train the Threshold Encoder which aim is to seperate the pred_map(the output of the network) to 1 or 0, but there is little differences between the pred_map after processed by the Threshold Encoder and the pred_map without processed by the Threshold Encoder.Details can look at the picture 微信图片_20211011153600 微信图片_20211011153606 微信图片_20211011153613

lab-gpu commented 2 years ago

Just supress a little noise in the background.Theses noise may not influence the result of crowd counting.Pred_map is so pretty good that the TE Module play a trivial module

lab-gpu commented 2 years ago

受不了了,说中文吧,你看第一张图片,预测到pred_map其实绝大部分,在目标区域已经是1,背景区域已经是0了,这是因为pred_map训练的时候,已经和mask_gt做了训练,那么训练出来的pred_map其实和mask_gt已经很接近了,我觉得通过TE(Threshold Encoder)模块,抑制掉的背景,其实真真真很少啊,不过可能这就是最后精度提升的原因吧

taohan10200 commented 2 years ago

你好,TE的主要目的不是去除背景,就像你说的,如果训练得好,背景区域基本已经没有噪声了,这种场景自然不需要TE。TE最主要的是想减少粘连的区域,例如,对于密集粘连区域,我们希望学出来的的阈值能够自适应的在这一块区域变大来实现更好的分割。