Research map / Detection and segmentation / CNN features
Fine-Grained Prototypes Distillation for Few-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 FPD, a meta-learning few-shot object detector on Meta R-CNN that distills support features into fine-grained prototypes (FFA) and adds balanced class-agnostic sampling (B-CAS) and a non-linear fusion module (NLF), evaluated on PASCAL VOC and MS COCO.
Problem
Few-shot object detection with meta-learning: class-level prototypes from the support branch are hard to generate precisely and lack detailed local information, and existing feature fusion is either dense/expensive or too simple for high-level relations.
Approach
Use per-class learnable feature queries with cross-attention to distill the support mid-level feature map into a small set of fine-grained prototypes, assign them to query features via attention with background prototypes, and pair this with balanced positive/negative class-agnostic sampling and a non-linear fusion head for high-level RoI-prototype fusion.
Key results
- On PASCAL VOC Novel Set 3 at 10 shots, FPD reaches 62.9 AP50 in single-run results, the best in that setting per the table.
- On MS COCO 30-shot, FPD achieves 19.3 AP averaged over multiple runs, tied for best among FR-CNN-based methods in the table (VFA 18.4).
- Ablation on VOC Novel Set 1: B-CAS plus NLF lifts 3/5/10-shot AP50 from a baseline of 56.7/58.3/61.4 to 62.8/67.1/66.3, and adding FFA reaches 64.0/67.6/68.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FPD instead of VFA | Meta-learning FSOD on PASCAL VOC Novel Set 1, 10-shot · PASCAL VOC Novel Set 1, 10-shot | AP50: 67.4 -> 68.4 | Helped |
| Use FPD instead of VFA | Meta-learning FSOD on PASCAL VOC Novel Set 2, 10-shot · PASCAL VOC Novel Set 2, 10-shot | AP50: 51.6 -> 53.9 | Helped |
| Use FPD instead of VFA | Meta-learning FSOD on PASCAL VOC Novel Set 3, 10-shot · PASCAL VOC Novel Set 3, 10-shot | AP50: 58.9 -> 62.9 | Helped |
| Use FPD instead of VFA | Meta-learning FSOD on MS COCO · MS COCO, 10-shot | AP: 15.9 -> 15.9 | No clear effect |
| Use FPD instead of VFA | Meta-learning FSOD on MS COCO · MS COCO, 30-shot | AP: 18.4 -> 19.3 | Helped |
| Add B-CAS (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 3-shot | AP50: 56.7 -> 61.2 | Helped |
| Add NLF to B-CAS (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 3-shot | AP50: 61.2 -> 62.8 | Helped |
| Add FFA to B-CAS and NLF (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 3-shot | AP50: 62.8 -> 64.0 | Helped |
| Add B-CAS (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 5-shot | AP50: 58.3 -> 64.7 | Helped |
| Add NLF to B-CAS (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 5-shot | AP50: 64.7 -> 67.1 | Helped |
| Add FFA to B-CAS and NLF (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 5-shot | AP50: 67.1 -> 67.6 | Helped |
| Add B-CAS (ablation) | Meta R-CNN on PASCAL VOC Novel Set 1 · PASCAL VOC Novel Set 1, 10-shot | AP50: 61.4 -> 64.9 | Helped |
Surprising findings
- Matching loss from prior work hurt performance when combined with FFA, attributed to asymmetry between the two branches.
- FPD underperforms VFA at 1- and 2-shot on VOC, which the authors attribute to difficulty capturing common features across very few shots.
Limitations
- Underperforms VFA in 1- and 2-shot VOC settings
- Gains on MS COCO are modest and the method ranks fourth among all methods
- Revised results in appendix show some reported numbers were erroneous in the original version
Experimental setup
Datasets: PASCAL VOC MS COCO
Benchmarks: PASCAL VOC MS COCO
Metrics: AP50 AP
Baselines: FSRW Meta R-CNN TFA w/ cos MPSR FSCE Meta FR-CNN Meta-DETR FCT VFA DeFRCN DCNet FSDetView
Builds on: Meta R-CNN DCNet VFA DETR TFA
Closest papers by meaning
- Frustratingly Simple Few-Shot Object Detection (ICML 2020 · similarity 0.92)
- Few-Shot Object Detection with Foundation Models (CVPR 2024 · similarity 0.89)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.89)
- Weakly Supervised Few-Shot Object Detection with DETR (AAAI 2024 · similarity 0.88)
- Domain-RAG: Retrieval-Guided Compositional Image Generation for Cross-Domain Few-Shot Object Detection (NeurIPS 2025 · similarity 0.88)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · 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.