Research map / Detection and segmentation / CNN features
MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers
Check 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 |
|---|---|---|---|
| directly predict class-labeled masks with a mask transformer and PQ-style training | MaX-DeepLab-L · COCO test-dev | PQ: 51.3 | Helped |
| replace convolutional blocks with axial-attention blocks (ablation) | MaX-DeepLab · COCO val | PQ: 45.7 -> 47.8 | Helped |
| increase input resolution from 641 to 1025 while using axial-attention blocks (ablation) | MaX-DeepLab · COCO val | PQ: 47.8 -> 49.4 | Helped |
| stack the decoder once (ablation) | MaX-DeepLab · COCO val | PQ: 1.4% PQ | Helped |
| extend decoder stacking from L=1 to L=2 (ablation) | MaX-DeepLab · COCO val | PQ: 47.1 -> 47.5 | No clear effect |
| train for more iterations (ablation) | MaX-DeepLab · COCO val | PQ: 45.7 -> 46.9 -> 47.7 | Helped |
| add P2M feedback attention (ablation) | MaX-DeepLab dual-path transformer · COCO val | PQ: 45.0 -> 45.7 | Helped |
| add M2M self-attention (ablation) | MaX-DeepLab dual-path transformer · COCO val | PQ: 45.1 -> 45.7 | No clear effect |
| apply transformer blocks at stride 8 in addition to stride 16 (ablation) | MaX-DeepLab dual-path transformer · COCO val | PQ: +1.0% PQ | Helped |
| add stride 4 transformer blocks to the stride 16 and 8 configuration (ablation) | MaX-DeepLab dual-path transformer · COCO val | PQ: 46.7 -> 46.3 | Hurt |
| use RQ × SQ rather than RQ + SQ for matching and loss weighting (ablation) | MaX-DeepLab · COCO val | PQ: 44.9 -> 45.7 | Helped |
| add instance discrimination loss to the PQ-style loss (ablation) | MaX-DeepLab · COCO val | PQ: 39.5 -> 42.6 | Helped |
Closest papers by meaning
- ReMaX: Relaxing for Better Training on Efficient Panoptic Segmentation (NeurIPS 2023 · similarity 0.90)
- Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation (CVPR 2020 · similarity 0.90)
- CMT-DeepLab: Clustering Mask Transformers for Panoptic Segmentation (CVPR 2022 · similarity 0.89)
- VIP-DeepLab: Learning Visual Perception With Depth-Aware Video Panoptic Segmentation (CVPR 2021 · similarity 0.88)
- MP-Former: Mask-Piloted Transformer for Image Segmentation (CVPR 2023 · similarity 0.87)
- PEM: Prototype-based Efficient MaskFormer for Image Segmentation (CVPR 2024 · similarity 0.86)
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.