[RE] FDA: Fourier Domain Adaptation for Semantic Segmentation

Arnesh Kumar Issar, Aryan Mehta, Karan Uppal, Saurabh Mishra, Kirtan Mali

Keywords: Semantic Segmentation, Unsupervised Domain Adaptation, Fourier Transform

Abstract:

The following submission is a reproducibility report for FDA: Fourier Domain Adaptation for Semantic Segmentation published in the CVPR 2020 as part of the ML Reproducibility Challenge 2020. We used the publicly available source code provided by the authors. Minor changes were made to the source code in order to load the model weights properly. The reproducibility experiments followed the training protocol as described in the original paper.

The paper describes a method for unsupervised domain adaptation, whereby the discrepancy between the source and target distributions is reduced by swapping the low-frequency spectrum of one with the other. The method is illustrated in the semantic segmentation domain. The method relies on Fourier Transform and its inverse instead of the model training for model alignment.

To formulate the report, extensive experimentation on the author's code has been conducted to verify all the claims made by the author described in detail below. We also performed additional experiments to gauge the extent of improvement brought upon by the method over the state-of-the-art methods.

Improvements to the code stack were made wherever for efficient execution on low computation systems.

Paper Url: https://openreview.net/forum?id=1LuPv8dPSZP&referrer=%5BML%20Reproducibility%20Challenge%202020%5D(%2Fgroup%3Fid%3DML_Reproducibility_Challenge%2F2020)