Great computer vision paper must have a cat in it

Task and data

At the start of this project (around May 2020), I start with Kaggle's Cat vs Dog to validate the basic idea. Basically, it's a binary classification task to tell an image is cat or dog. And we collected a lot of eye gaze from our lab members.

The collected data is then used to train a network. This exploration did not reach the pushlish quality, but I believe it is worth to mention.

Result

The image below is trained with 1000 training images.

Gotta love these cute little guys

And we try more class labels, then we found:

for the network attetion (CAM), more label=less label+gaze.

You may notice we did not report the classification accuracy, because this problem is so easy for these networks, 95%+ acc can be easily achieved.