Closed ccJia closed 6 years ago
Thank you for your question! It is a common phenomenon that the network tends to be affected by majority labels in classification. To avoid the effect of it we used a weighted sigmoid cross entropy loss, following previous work. For example, if the positive ratio is too large for attribute "trousers", we lower that loss penalty for positive samples and use higher loss weight for negative attributes. The formulation of the loss function can be found in the paper.
Thank you for your reply. I'll try this method .^-^
@ccJia hey, dude, are you re implementing the paper in Caffe? lol:)
@JohnHush Yes , I am implementing the paper. And I just finished the attribute extracting . I want add the attention net in it .
@ccJia me, too. I'm wondering how to add attention in the M-NET, need to build new layer?? do you ha ve some suggestion?
@JohnHush We stand on the same starting line. I just asked my friend about this. ^-^ If there is any good idea, I will contact you .
@ccJia that's cool, recently I've implemented DeepMAR, I supposed it's basically the same as M-NET without Attention, keep in touch
Hi, have you re-implemented this paper? I am doubt about how to make use of attention. Can you give some suggestions? Thank you. @ccJia
@JohnHush Hi, may you have already reproduced DeepMar? I have been studying it recently. Can you share your results?
@kclch ,hi, yes i have reproduced DeepMar, just follow the instruction in the paper"Multi-attribute Learning for Pedestrian Attribute Recognition in Surveillance Scenarios" . And you could modify the base net by yourself, like transfer to ResNet or something else, and replace the last layer to multi task layer, and I have used the dataset published by @xh-liu in this paper, so grateful. I hope it helps
@JohnHush Can you share your entire program with me? I will be very grateful to you.
@kclch i've implemented in caffe, you could look into this directory:
And I can't find the "multilabel.prototxt" file
@JohnHush Can you provide the test code, thank you
Hi Liu,
I tried to duplicate your paper, but I met some questions about the attribute recognition. My net is composed of "res-net" and a "SigmoidCrossEntropyLoss" layer. And the attributes "trousers" and "Age 18-60" occupy a big proportion in "PA-100K" dataset. And it would lead a phenomenon that when I enter a part of a picture the net could also give the labels include "trousers" and "Age 18-60"(I just want a single attribute ). Could you give me some advice to eliminate these distractions? Thank you.