Research map / Detection and segmentation / Open-set and few-shot detection
Learning Mask-aware CLIP Representations for Zero-Shot Segmentation
Siyu Jiao, Yunchao Wei, Yaowei Wang, Yao Zhao, Humphrey Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 MAFT to FreeSeg | FreeSeg with ViT-B/16 CLIP · COCO-Stuff, Pascal-VOC, and ADE20K zero-shot | mIoU unseen: 42.2 -> 50.4; 78.6 -> 81.8; 4.4 -> 8.7 | Improved |
| Add MAFT to FreeSeg | FreeSeg with ViT-B/16 CLIP · A-847, A-150, PC-459, PC-59, and PAS-20 | mIoU: +3.0%; +11.2%; +6.4%; +19.1%; +4.4% | Improved |
| Add IP-CLIP Encoder (ablation) | FreeSeg · COCO-Stuff (no ensemble) | mIoU unseen: 29.3 -> 36.2 | Improved |
| Add mask-aware fine-tuning loss (ablation) | FreeSeg + IP-CLIP · COCO-Stuff (no ensemble) | mIoU unseen: 36.2 -> 47.1 | Improved |
| Add self-distillation loss (ablation) | FreeSeg + mask-aware fine-tuning · COCO-Stuff (no ensemble) | mIoU unseen: 47.1 -> 49.7 | Improved |
| Use SmoothL1 instead of KL for the mask-aware loss (ablation) | MAFT without self-distillation · COCO-Stuff (no ensemble) | mIoU unseen: 41.8 -> 47.1 | Improved |
| Increase fine-tuning from 1k to 2k, 3k, 4k, or 5k iterations (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 49.7 -> 47.6; 49.7 -> 46.5; 49.7 -> 46.1; 49.7 -> 45.7 | Hurt |
| Freeze the class embedding or positional embedding (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 44.7; 44.9 | Hurt |
| Freeze the Transformer MLP (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 44.7 -> 48.7 | Improved |
| Freeze convolution layers (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 44.7 -> 49.7 | Improved |
| Start mask attention at layer 8 rather than layer 0 (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 46.4 -> 49.7 | Improved |
| Start mask attention at layer 10 rather than layer 8 (ablation) | MAFT · COCO-Stuff (no ensemble) | mIoU unseen: 49.7 -> 45.7 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FreeSeg — FreeSeg: Unified, Universal and Open-Vocabulary Image Segmentation (CVPR 2023) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.89)
- Spectral Prompt Tuning: Unveiling Unseen Classes for Zero-Shot Semantic Segmentation (AAAI 2024 · similarity 0.89)
- Exploring Regional Clues in CLIP for Zero-Shot Semantic Segmentation (CVPR 2024 · similarity 0.88)
- DiffCut: Catalyzing Zero-Shot Semantic Segmentation with Diffusion Features and Recursive Normalized Cut (NeurIPS 2024 · similarity 0.88)
- Building Vision-Language Models on Solid Foundations with Masked Distillation (CVPR 2024 · similarity 0.88)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · 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.
