Xiaolin Zhang

Hi， We apply two classifiers. The final loss value is computed by adding the cross-entropy loss of them. We do not apply any reverse functions.

The feature maps to be erased are from the activation function ReLU, whose range is [0, +inf). Therefore, the actual erasing operation is to set the pixel values of object...

Hi Rahul, It is definitely feasible to train the model on your custom dataset. You need to replace the training list in datalist and then modifying the bash scripts. This...

I use the code from CAM(https://github.com/metalbubble/CAM) for producing boxes and measuring localization accuracy.

> Hi @xiaomengyc , > > Thanks for sharing your code! As I am running your code, I get a similar classification error on CUB dataset. However, when it comes...

Hi @chaoyan1037 , Sorry for the missing threshold values. I looked through my experiment records but still did not find these values. It seems I only picked out the best...

@xiaomengyc ACoL is exactly designed for the weakly supervised localization problem. It actually learns a classification network with the supervision of image level labels (like cat or dog). We apply...

@tim20120526 Images are firstly resized to 256*256, and then cropped to 224*224 in the training stage. For the testing stage, images are directly resized to 224*224. Sorry, I didn't get...

@laukun Sorry, I did not compare the loss curves between the two methods. Although these two methods have the same mathematical formalization, the network structures are not exactly the same....

The lists of CUB and its training scripts are uploaded.