Research map / Detection and segmentation / Open-set and few-shot detection
Frustratingly Simple Few-Shot Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| fine-tune only the box predictor while freezing the feature extractor | Faster R-CNN · PASCAL VOC Novel Set 1, 1-shot | AP50: 19.9 -> 39.8 | Helped |
| fine-tune only the box predictor while freezing the feature extractor | Faster R-CNN · COCO novel classes, 10-shot | AP: 8.7 -> 10.0 | Helped |
| use a cosine-similarity box classifier instead of a fully connected classifier (ablation) | TFA · PASCAL VOC Novel Set 1, 1-shot | AP50: 36.8 -> 39.8 | Helped |
| use a cosine-similarity box classifier instead of a fully connected classifier (ablation) | TFA · PASCAL VOC Novel Set 1, 10-shot | AP50: 57.0 -> 56.0 | No clear effect |
| initialize the novel classifier randomly instead of using fine-tuned novel weights (ablation) | TFA · PASCAL VOC split 3, 1-shot | Novel AP: 13.4 -> 15.6 | Helped |
| initialize the novel classifier from fine-tuned novel weights instead of random weights (ablation) | TFA · COCO, 1-shot | Novel AP: 3.2 -> 3.4 | Helped |
| set cosine-similarity scale to 20 rather than 10 (ablation) | TFA cosine classifier · PASCAL VOC split 3, 1-shot | Novel AP: 9.9 -> 15.6 | Helped |
| set cosine-similarity scale to 20 rather than 10 (ablation) | TFA cosine classifier · COCO, 10-shot | Novel AP: 4.7 -> 10.0 | Helped |
| evaluate using repeated random samples and report their mean instead of relying on one sample | few-shot object detection benchmarks · PASCAL VOC, 1-shot, split 1 | AP50: around 15 points lower | Hurt |
| use the two-stage fine-tuning scheme without repeated sampling | LVIS · rare classes | AP: around 13 points | Helped |
Closest papers by meaning
- Meta-Tuning Loss Functions and Data Augmentation for Few-Shot Object Detection (CVPR 2023 · similarity 0.94)
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.92)
- Incremental-DETR: Incremental Few-Shot Object Detection via Self-Supervised Learning (AAAI 2023 · similarity 0.90)
- A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps (CVPR 2026 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.89)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.88)
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.