How to interpret Class Activation Maps

Deep Learning
1

Hi guys!

To give some context, I have been experimenting with fashion images using the deepfashion dataset and I was interested in learning about classification using CNNs, so I built a resnet-50 classifier and it was performing okayish.

So, I started looking into what could use in order to understand where is the model getting confused the most and as mentioned in this thread, I used Class Activation Maps to visualise the results of my prediction but I cannot seem to find much help with this.

For eg. I am trying to look at whether a dress has an off-shoulder, a cold-shoulder, a no-shoulder-type or a one-shoulder type. I built a single-label 4 class classifier and used the CAM technique to overlay the map over those images where the model was primarily making mistakes.

Now, since the shoulder type is what I am trying to predict I could see that the neck-shoulder region lights up in the images but along with it, many other regions light up as shown below. And sometimes, the model predicts the wrong attribute with a very high confidence i.e. probability corresponding to predicted label is more than 80%.

417279
417279720Ă—720 44.6 KB

414588
414588720Ă—720 49.9 KB

In the above figures, T corresponds to the True Label and P corresponds to the predicted label. The training data is balanced with almost 800 images per class i.e. 3200 images in total trained using Adam optimizer for 21 epochs with a final validation accuracy of close to 85%. However when investigating the images where model is incorrectly predicting the tags, I am yet not able to figure it out. Also, in what other ways could CAM be used for interpreting the obtained results?

Any help would be much appreciated.

Thanks!

2 Likes
2

Hi @ElisonSherton,

I’m not too sure why those areas are being lit up either, but here’s an idea I had:

  1. In both images the model is paying attention to the person’s feet. Maybe there is a correlation between someone’s footwear and their dress type in the dataset. For example, if 90% of the people wearing dark shoes (like the bottom photo) are also wearing cold-shoulder dresses, the model might learn to always predict cold-shoulder whenever it sees dark shoes.

Also, are the labels correct? I don’t know much about dress shoulder types, but those 2 dresses look pretty different in the shoulders but they’re both labeled as “no-shoulder type”.

Let me know if either of those ideas end up helpful!

3

Thanks @GoofyMango. This was helpful. There were also other useful details which we could get thanks to your observation.

2 Likes