Summary of Lightweight Frequency Masker For Cross-domain Few-shot Semantic Segmentation, by Jintao Tong et al.
Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation
by Jintao Tong, Yixiong Zou, Yuhua Li, Ruixuan Li
First submitted to arxiv on: 29 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a new approach to cross-domain few-shot segmentation (CD-FSS), which pre-trains a model on a large-scale source-domain dataset and then transfers it to data-scarce target-domain datasets for pixel-level segmentation. The authors identify an intriguing phenomenon: filtering different frequency components for target domains can significantly improve performance, often by as much as 14% mIoU. They delve into this phenomenon to interpret the results and find that the reduced inter-channel correlation in feature maps enhances robustness against domain gaps and larger activated regions for segmentation. Building on this insight, they propose a lightweight frequency masker with an Amplitude-Phase Masker (APM) module and an Adaptive Channel Phase Attention (ACPA) module. The APM module introduces only 0.01% additional parameters but improves average performance by over 10%, while the ACPA module imports only 2.5% parameters and further improves performance by over 1.5%, surpassing state-of-the-art CD-FSS methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Cross-domain few-shot segmentation is a new approach to pixel-level segmentation that pre-trains models on large datasets and then transfers them to smaller datasets for segmentation. The authors found that filtering different frequency components in the target dataset can make the model work much better, sometimes by as much as 14%. They tried to understand why this happens and found that it’s because the model is less likely to get confused between different features when the data is filtered in this way. They also came up with a new way to do this filtering using two special modules that add only a few extra calculations to the model. This makes their approach better than other state-of-the-art methods. |
Keywords
» Artificial intelligence » Attention » Few shot
