9 October 2023
Computer Vision and Pattern Recognition
Batch norm updating is ineffective for segmentation TTA
Teacher-student stabilizes but does not improve segmentation TTA
Severe long-tail challenge in segmentation TTA
Test augmentation partially relieves long-tail issue
Classic TTA methods fail for segmentation tasks
Classic test-time adaptation methods fail for semantic segmentation
This paper systematically investigates classic test-time adaptation (TTA) methods for semantic segmentation. Through extensive experiments, it finds that techniques effective for classification TTA, like batch norm updating and teacher-student schemes, do not work well for segmentation. Key challenges are inaccurate distribution estimation and long-tailed class imbalance. The paper provides insights to guide future segmentation TTA research.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper