Research map / Detection and segmentation / Open-set and few-shot detection
Exploring Base-Class Suppression with Prior Guidance for Bias-Free One-Shot Object Detection
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 BSPG, a one-shot object detector that suppresses base-class distractors with a base-class predictor and base-class suppression module, plus a non-parametric prior guidance map from high-level feature correlation.
Problem
One-shot object detection suffers from model bias toward base classes (false positives on base-class objects) and degraded generalization to novel classes trained with abundant base data.
Approach
Adds a base-class predictor whose confident base detections are used to learnably suppress novel-class scores via a base-class suppression module, and a parameter-free prior guidance map computed from cosine similarity of frozen high-level features to reweight target features.
Key results
- On COCO OSOD (AP50), novel-class average improves from 25.6 (BHRL) to 27.3, and base-class average from 53.5 to 55.0.
- On COCO FSOD 1-shot, achieves 15.5 AP vs. 11.9 for DAnA (2-classification baseline), with 3.5 AP50 gain.
- Ablation on COCO split-2: BcS alone adds 1.5 AP50 and PG adds 0.9 AP50; combined gives 30.7 AP50 vs. 28.5 baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add base-class suppression (BcS) module (ablation) | BSPG novel-class predictor · COCO split-2 novel classes, 1-shot | AP50: +1.5 AP50 (28.5 to 30.0) | Helped |
| add prior guidance (PG) module (ablation) | BSPG novel-class predictor · COCO split-2 novel classes, 1-shot | AP50: +0.9 AP50 (28.5 to 29.4) | Helped |
| add both BcS and PG together (ablation) | BSPG · COCO split-2 novel classes, 1-shot | AP50: +2.2 AP50 over baseline (28.5 to 30.7) | Helped |
| set lambda1=lambda2=0.5 vs. supervising only the final output (lambda1=1, lambda2=0 gives 27.8) (ablation) | BSPG loss · COCO split-2 novel classes, 1-shot | AP50: 30.7 vs 27.8 (lambda2=0) | Helped |
| BSPG vs BHRL on COCO | OSOD · COCO novel and base classes | AP50: +1.9 novel, +1.4 base | Helped |
| BSPG vs BHRL on PASCAL VOC | OSOD · PASCAL VOC novel classes | AP50: +0.9 novel, +0.5 base | Helped |
| BSPG vs DAnA on COCO few-shot | FSOD · COCO novel classes, 1/3/5-shot | AP50: +3.5 (1-shot), +4.6 (3-shot), +5.0 (5-shot) | Helped |
Surprising findings
- Gains on PASCAL VOC are smaller than on COCO because VOC images rarely contain base-class distractors.
Limitations
- Base-class predictor's proposals are mostly base-class, so its RPN is removed, adding a separate base-class detector to train.
- Gains on VOC are modest because scenes rarely contain base-class distractors.
Experimental setup
Datasets: COCO PASCAL VOC ImageNet (reduced)
Benchmarks: COCO PASCAL VOC
Metrics: AP50 AP AP75 precision
Baselines: CoAE AIT SaFT BHRL FSDetView Meta-DETR DeFRCN FGN Attention RPN DAnA
Builds on: Faster R-CNN BHRL Attention RPN FPN
Closest papers by meaning
- Few-Shot Object Detection via Association and DIscrimination (NeurIPS 2021 · similarity 0.87)
- Adaptive Image Transformer for One-Shot Object Detection (CVPR 2021 · similarity 0.85)
- Don't Even Look Once: Synthesizing Features for Zero-Shot Detection (CVPR 2020 · similarity 0.85)
- Feature Selective Anchor-Free Module for Single-Shot Object Detection (CVPR 2019 · similarity 0.85)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.84)
- UN-DETR: Promoting Objectness Learning via Joint Supervision for Unknown Object Detection (AAAI 2025 · 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.