Research map / Detection and segmentation / Open-set and few-shot detection
Generalized Class Discovery in Instance Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a generalized class discovery method for instance segmentation with instance-wise temperature assignment, class-wise dynamic reliability criteria for pseudo-labels, and a soft attention module, evaluated on COCO_half+LVIS and LVIS+VG.
Problem
Generalized class discovery in instance segmentation on long-tailed datasets, where instance distributions per class are imbalanced and pseudo-labels for tail classes are often excluded by fixed reliability criteria.
Approach
Assigns per-instance contrastive temperatures based on estimated headness (head samples get relaxed discrimination), uses class-wise reliability thresholds that are dynamically tightened over training to filter pseudo-labels, and adds a pooled-feature soft attention module trained with boundary-weighted masks.
Key results
- On COCO_half + LVIS, the method reaches 12.85 mAP_all, versus 9.42 for the best prior baseline NCDLR.
- On LVIS + VG, the method reaches 5.21 mAP_all, versus 4.71 for NCDLR.
- Ablation: ITA adds about 1.85 mAP_all over baseline and SAM and RDL add further gains to 12.85.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add instance-wise temperature assignment (ablation) | GCD model · COCO_half + LVIS | mAP_all: 5.79 -> 7.64 | Helped |
| Add soft attention module after adding ITA (ablation) | GCD model · COCO_half + LVIS | mAP_all: 7.64 -> 10.71 | Helped |
| Add reliability-based dynamic learning after adding ITA and SAM (ablation) | Instance segmentation training · COCO_half + LVIS | mAP_all: 10.71 -> 12.85 | Helped |
| Use ITA instead of temperature scheduling (ablation) | GCD model · COCO_half + LVIS; τmin 0.07, τmax 0.5 | mAP_all: 10.86 -> 11.67 | Helped |
| Use ITA instead of temperature scheduling (ablation) | GCD model · COCO_half + LVIS; τmin 0.07, τmax 1 | mAP_all: 11.43 -> 12.85 | Helped |
| Set τmax to 1 rather than 0.5 in ITA (ablation) | GCD model · COCO_half + LVIS | mAP_all: 11.67 -> 12.85 | Helped |
| Use SAM instead of RGA-S (ablation) | GCD model · COCO_half + LVIS | mAP_all: 11.62 -> 12.85 | Helped |
| Use class-wise dynamic reliability criteria instead of fixed global criteria (ablation) | Pseudo-label training · COCO_half + LVIS | mAP_all: 11.79 -> 12.85 | Helped |
| Use GGN for class-agnostic instance masks instead of UDOS (ablation) | Proposed method · COCO_half + LVIS | mAP_all: 12.54 -> 12.85 | Helped |
| Use the proposed GCD model instead of NCDLR | Proposed method · COCO_half + LVIS | mAP_all: 9.42 -> 12.85 | Helped |
| Use the proposed GCD model instead of NCDLR | Proposed method · LVIS + VG | mAP_all: 4.71 -> 5.21 | Helped |
Limitations
- Assumes full availability of labeled and unlabeled data from the start, so it is unsuitable for sequentially arriving data.
- Assumes prior knowledge of the total number of classes.
Experimental setup
Datasets: LVIS COCO Visual Genome ImageNet
Benchmarks: LVIS Visual Genome
Metrics: mAP mAP_all mAP_known mAP_novel
Baselines: k-means RNCDL ORCA UNO SimGCD μGCD NCDLR TS
Builds on: GCD (Vaze et al.) SimGCD ORCA SOLOv2 ST++ RNCDL
Closest papers by meaning
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.88)
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.86)
- DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (ICLR 2025 · similarity 0.86)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.86)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.85)
- Adaptive Part Learning for Fine-Grained Generalized Category Discovery: A Plug-and-Play Enhancement (CVPR 2025 · similarity 0.85)
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.