Research map / Detection and segmentation / CNN features
CLUSTSEG: Clustering for Universal 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 |
|---|---|---|---|
| Use task-specific Dreamy-Start initialization and Recurrent Cross-Attention instead of learned queries and standard cross-attention (ablation) | ClustSeg · COCO Panoptic val | PQ: 49.7 -> 54.3 | Helped |
| Add Dreamy-Start query initialization (ablation) | COCO Panoptic baseline · COCO Panoptic val | PQ: 49.7 -> 51.0 | Helped |
| Add Recurrent Cross-Attention (ablation) | COCO Panoptic baseline · COCO Panoptic val | PQ: +3.5% PQ | Helped |
| Use scene-adaptive initialization for thing centers (ablation) | Dreamy-Start · COCO Panoptic val | PQ^Th: +1.1% | Helped |
| Use scene-agnostic initialization for stuff centers (ablation) | Dreamy-Start · COCO Panoptic val | PQ^St: 44.9% -> 45.7% | Helped |
| Use scene-adaptive initialization for stuff centers (ablation) | Dreamy-Start · COCO Panoptic val | PQ^St | No clear effect |
| Replace recurrent cross-attention with vanilla cross-attention (ablation) | ClustSeg · COCO Panoptic val | PQ: 51.0 -> 54.3 | Hurt |
| Replace recurrent cross-attention with K-Means cross-attention (ablation) | ClustSeg · COCO Panoptic val | PQ: 53.4 -> 54.3 | Hurt |
| Increase recurrent clustering iterations from 1 to 3 (ablation) | ClustSeg · COCO Panoptic val | PQ: 53.8 -> 54.3 | Helped |
| Increase recurrent clustering iterations beyond 3 (ablation) | ClustSeg · COCO Panoptic val | PQ: 54.3 -> 54.3 | No clear effect |
| Replace recurrent updates with multiple non-recurrent cross-attention layers (ablation) | ClustSeg · COCO Panoptic val | PQ: 54.3 -> 53.9 | Hurt |
| Remove deep supervision from intermediate E-steps (ablation) | ClustSeg · COCO Panoptic val | PQ: 53.0 -> 54.3 | Hurt |
Closest papers by meaning
- ClusterFomer: Clustering As A Universal Visual Learner (NeurIPS 2023 · similarity 0.89)
- Unsupervised Universal Image Segmentation (CVPR 2024 · similarity 0.85)
- Recurrent Pixel Embedding for Instance Grouping (CVPR 2018 · similarity 0.85)
- FasterSeg: Searching for Faster Real-time Semantic Segmentation (ICLR 2020 · similarity 0.85)
- Image Clustering Conditioned on Text Criteria (ICLR 2024 · similarity 0.84)
- Segment Anything without Supervision (NeurIPS 2024 · similarity 0.84)
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.