NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3479
Title:Grid Saliency for Context Explanations of Semantic Segmentation

Reviewer 1

1. Grid saliency can produce spatially coherent explanations for dense prediction networks firstly. This work seems to be sufficiently innovative and meaningful. 2. Experiment results are clear in the context biases research on synthesis dataset. However, a comparative analysis should be more detailed. The grid saliency is based on perturbation saliency method, while other contrast methods are belong to gradient-based saliency methods. I have doubts about the relationship between these methods. ------------------------------------ The author responses provided a comparative analysis of two methods. 3. For the real-world dataset cityscapes, more quantitative results are expected to be provided in the submission like the numerical results in supplementary material. I am also interested in how to use context explanations of erroneous predictions to improve the performance. --------------------------- The author responses provided quantitative experiment results and more details about the practice.

Reviewer 2

1. The major concern with this works is that the impact of the work seems to be limited, and I am not sure if the practitioners would be interested in grid saliency and how would this be useful for them. 2. The explanation of how to obtain grid saliency map is not clear. For example, in equation 2, the M*_grid is not clear how to obtain from the optimization. How would the size of R affect the M? Would we always be able to obtain an optimal M*_grid given any request mask R? It is also not very clear how to twist the lambda to make this works.

Reviewer 3

ORIGINALITY To the best of my knowledge, there has been no previous study on deep neural networks about explaining pixel-wise predictions in image segmentation. QUALITY The reported results indicate the superiority of this perturbation-based method with respect to three gradient-based methods. Results seem convincing. CLARITY The text is clearly written, with the necessary visualizations and results, all of them commented to guide the reader's understanding. SIGNIFICANCE TRhe work makes an interesting and necessary step that will help the community working on this task, and any other pixel-wise one.