peternara/rectified-gradient-training-saliency-segmentation

Official repository for "Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps".

Rectified Gradient

Experiment codes and additional samples for Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps.

The work is presented at 2019 ICCV Workshop on VXAI.

Random Sample Links

We list the links to Google drives containing attribution maps for 1.5k randomly chosen ImageNet images. Samples for the same image have the same file name {image number}.png for ease of comparison between attribution maps with and without final threshold / attribution maps for original image and adversarial image. We did not include samples for adversarial attacks which failed to change the final decision.

Random Samples W/O Baseline Final Threshold

Random Samles with Baseline Final Threshold

Adversarial Attack Samples W/O Baseline Final Threshold

Adversarial Attack Samples with Baseline Final Threshold

Results

Effect of Threshold Percentile

Comparison with Other Attribution Methods