Research map / Detection and segmentation / CNN features
ReMaX: Relaxing for Better Training on Efficient Panoptic Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ReMaX to kMaX-DeepLab under 50K-iteration schedule | kMaX-DeepLab R50 · COCO panoptic val | PQ: +3.6 PQ | Helped |
| Train ReMaX longer (200K iterations) | ReMaX R50 · COCO panoptic val | PQ: 52.4 -> 54.2 | Helped |
| Add ReMaX to MobileNetV3-Small/Large at 50K iterations | kMaX-DeepLab · COCO panoptic val | PQ: +4.9 and +5.2 PQ | Helped |
| Add ReMaX to MobileNetV3-Small/Large at 150K iterations | kMaX-DeepLab · COCO panoptic val | PQ: +3.3 and +2.5 PQ | Helped |
| Use sigmoid instead of softmax activation (ablation) | ReMaX · COCO panoptic val | PQ: 49.5 -> 52.4 | Helped |
| Add ReMaX (sigmoid) in place of gradient clipping (ablation) | sigmoid baseline · COCO panoptic val | PQ: 51.2 -> 52.4 | Helped |
| Add gradient clipping to sigmoid baseline (ablation) | sigmoid baseline · COCO panoptic val | PQ: 50.4 -> 51.2 | Helped |
| Increase number of ReMasks from 4 to 6 (ablation) | ReMaX · COCO panoptic val | PQ: 52.4 -> 51.5 | Hurt |
| Increase number of ReMasks from 0 to 4 (ablation) | ReMaX · COCO panoptic val | PQ: 50.4 -> 52.4 | Helped |
| Set ReClass eta to 0.1 versus 0 (ablation) | ReMaX · COCO panoptic val | PQ: 51.7 -> 52.4 | Helped |
| Increase ReClass eta from 0.1 to 0.2 (ablation) | ReMaX · COCO panoptic val | PQ: 52.4 -> 51.5 | Hurt |
| Remove aux semantic head at test time (with identity mapping) (ablation) | ReMaX · COCO panoptic val | PQ: 52.4 -> 52.4 | No clear effect |
Closest papers by meaning
- MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers (CVPR 2021 · similarity 0.90)
- MP-Former: Mask-Piloted Transformer for Image Segmentation (CVPR 2023 · similarity 0.89)
- UniDAformer: Unified Domain Adaptive Panoptic Segmentation Transformer via Hierarchical Mask Calibration (CVPR 2023 · similarity 0.88)
- Unifying Training and Inference for Panoptic Segmentation (CVPR 2020 · similarity 0.87)
- PanopticDepth: A Unified Framework for Depth-Aware Panoptic Segmentation (CVPR 2022 · similarity 0.87)
- Real-Time Panoptic Segmentation From Dense Detections (CVPR 2020 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.