Gradcam highlighting activation region on upper-left corner

[olguin]

First of all, thank you very much for you support and great work. To Clarify my background, I have lot of experience in software engineering but only a few months experience on CV/ML.

We are trying to differentiate apart clockwise vs counter clockwise hands movement by training a 2 class Classifier based on action recognition models like i3D, slowfast both pretrained on kinetics400 and later trained on our own custom dataset . Presently we have collected around 2300 samples combined for all 3 datasets (Train, validation, test).

in order to understand what the neural network is doing we have implemented gradcam using your script in demo/demo_gradcam.py, in the layer suggested in the doc: backbone/layer4/1/relu . The problem we are seeing is that very frequently (in the majority of our test videos) the upper left (top-left) corner of the videos is getting highlighted. To be clear, that's not the only region highlighted for the same video but different frames there are others, sometimes focusing on the hand, the hand knuckles, the screwdriver, or other part of the frame.

The following grad cam images, were gotten using i3d+resnet50 video (pretrained on kinetics400) and trained on our custom dataset:

Attaching here also the configuration file used:
clockwise_mov_i3D_resnet50_kinetics400.txt

We also witnessed the same upper-left highlighting behavior on another 2 class classifier we trained "opening or closing door vs putting or getting something from a table". Videos extracted from Charade dataset.

So Is this normal/expected on any gradcam for these models? or is there something wrong with our gradcam use or its implementation?

Any help will be greatly appreciated!!., thank you very much in advance,

Juan Pablo Olguin