Output of Cam doesn't match the bounding box

[sriniwaas]

@brucechou1983 : After training, using cam.py, I am getting the predicted cam output and bounding box mismatch. Please help.

[walterwang]

were u able to find a solution on this?

[jennyzu]

I have the same issue. Does anybody have a solution or idea?

[Sadam1195]

@jennyzu @walterwang I don't see any normalization (with imagenet_mean and imagenet_std values) in cam.py since transform_batch_images is not called. I think due to this test.py and cam.py return different predictions for same images hence the box doesn't fit right on heatmap. As I understand this normalization is a part of input pipeline during training, so this function should be called in cam.py, too. When you include transform_batch_images in cam.py output will be correct heatmap so you'll get correct bounding box.

[gmarceca]

In addition to @Sadam1195 comment I am afraid there is a bug here https://github.com/brucechou1983/CheXNet-Keras/blob/master/cam.py#L54 and class_weights[index] should be replaced by class_weights[:, index] in order to sum across all last conv layer filters (1024) for a given index. You can see an example here https://github.com/jacobgil/keras-cam/blob/master/cam.py#L34

[monteiro141]

In addition to @Sadam1195 comment I am afraid there is a bug here https://github.com/brucechou1983/CheXNet-Keras/blob/master/cam.py#L54 and class_weights[index] should be replaced by class_weights[:, index] in order to sum across all last conv layer filters (1024) for a given index. You can see an example here https://github.com/jacobgil/keras-cam/blob/master/cam.py#L34

By adding class_weights[:, index] my final grad cam is weird