Research map / Detection and segmentation / Open-set and few-shot detection
Reference Twice: A Simple and Unified Baseline for Few-Shot Instance Segmentation
Yue Han, Jiangning Zhang, Yabiao Wang, Chengjie Wang, Yong Liu, Lu Qi, Ming–Hsuan Yang, Xiangtai Li
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 query-based second reference module (ablation) | Mask2Former baseline · COCO minival K=10 | nAP: 14.7 -> 16.3 | Improved |
| add image-based first reference module on top of second reference (ablation) | RefT · COCO minival K=10 | nAP: +1.1 | Improved |
| second reference reduces base-class bias (ablation) | Mask2Former · COCO minival K=10 | bAP: +1.6 nAP, -0.5 bAP | Mixed |
| first reference helps both novel and base classes (ablation) | Mask2Former · COCO minival K=10 | nAP and bAP: +0.9 nAP, +0.3 bAP | Improved |
| replace cross-attention with adaptive convolution in second reference (ablation) | RefT · COCO gFSIS K=10 | nAP: 17.4 -> 16.8 | Hurt |
| add auxiliary classification loss in second reference (ablation) | RefT · COCO gFSIS K=10 | nAP: 17.4 -> 17.5 | No clear effect |
| fine-tune CAL during novel fine-tuning (vs frozen) (ablation) | RefT · COCO minival K=10 | nAP: 2.6 -> 17.4 | Improved |
| use larger-scale support features (Res3 vs Res5) in first reference (ablation) | RefT · COCO minival K=10 | nAP: 16.4 -> 17.2 | Improved |
| use query-level instead of feature-level enhancement in second reference (ablation) | RefT · COCO minival K=10 | nAP: +1.1 | Improved |
| add Class-Enhanced BKD Loss instead of L2 weight regularization (ablation) | RefT · COCO iFSIS K=10 | bAP: 29.4 -> 33.2 | Improved |
| use BKD loss instead of weight regularization (ablation) | RefT · COCO iFSIS K=10 | nAP: 11.5 -> 10.1 | Hurt |
| use Bayesian logit classifier instead of cosine classifier (ablation) | RefT · COCO iFSIS K=10 | nAP: +1.6 nAP, +0.8 bAP | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- iFS-RCNN — iFS-RCNN: An Incremental Few-Shot Instance Segmenter (CVPR 2022) · better on 3 of 4 shared results
- Meta-DETR — Meta-DETR: Image-Level Few-Shot Detection with Inter-Class Correlation Exploitation (TPAMI 2022) · better on 2 of 2 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
- Incremental Few-Shot Instance Segmentation (CVPR 2021 · similarity 0.90)
- iFS-RCNN: An Incremental Few-Shot Instance Segmenter (CVPR 2022 · similarity 0.89)
- FGN: Fully Guided Network for Few-Shot Instance Segmentation (CVPR 2020 · similarity 0.89)
- Less than Few: Self-Shot Video Instance Segmentation (ECCV 2022 · similarity 0.88)
- Foreground-Covering Prototype Generation and Matching for SAM-Aided Few-Shot Segmentation (AAAI 2025 · similarity 0.87)
- Beyond Pixel and Object: Part Feature as Reference for Few-Shot Video Object Segmentation (AAAI 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.
