Yi-Hsuan Tsai

Have you compiled matcaffe by "make matcaffe"?

@Ruofei-Li yes, you need the configuration file contained in this repo.

The first one is the initial response map (used as the pseudo ground truths later, Sec 4.2). The second one is the final semantic segmentation output after training the model...

y_20 is the output (x_20 is before sigmoid) of 20-category classification, and y_200 is the output of the fine-grained classification, where each parent class contains 10 sub-categories. And yes, the...

@TyroneLi @zerodohero The model architecture is the same. We did not heavily tune the model but just lower the learning rate a bit to account for noisy pseudo labels.

There is a `train_aug.txt` in the [script](https://github.com/Juliachang/SC-CAM/blob/master/script.sh#L24).

no, it's the typical classification network but with two heads: parent classifier and sub-category classifier.

What does your label look like, e.g., only [0, 1] categories? If so, please make sure that you have the correct size in the classifier (final output layer), so that...

Based on experiments we tried before, feature space adaptation is more sensitive and would need a much smaller weight for the adversarial loss.

Please make sure that the pre-trained weight is successfully loaded (it also happened to us that the performance could drop to 2x if not doing so). Normally, the model should...