Research map / Continual learning / Class-incremental
NIFF: Alleviating Forgetting in Generalized Few-Shot Object Detection via Neural Instance Feature Forging
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 |
|---|---|---|---|
| Replay forged base instance features during novel-data finetuning without storing base data | DeFRCN · MS-COCO 10-shot | AP: 18.2 -> 32.2 | Helped |
| Use NIFF instead of the strongest listed base-data-free baseline | DeFRCN · MS-COCO 10-shot | AP: 31.4 -> 32.2 | Helped |
| Combine separate class-aware heads with classwise statistics rather than separate heads with full-data statistics (ablation) | NIFF generator · MS-COCO 10-shot | AP: 29.0 -> 30.7 | Helped |
| Use classwise statistics with a shared generator head instead of full base-data statistics (ablation) | NIFF generator · MS-COCO 10-shot | AP: 29.3 -> 28.5 | Hurt |
| Use separate class-aware heads instead of a shared head with full base-data statistics (ablation) | NIFF generator · MS-COCO 10-shot | AP: 29.3 -> 29.0 | No clear effect |
| Add the generator cross-entropy term (ablation) | NIFF generator · MS-COCO 10-shot | AP: 30.0 -> 30.7 | Helped |
| Increase generator channels from 8 to 64 (ablation) | NIFF generator · MS-COCO 10-shot | AP: 30.7 -> 30.7 | No clear effect |
| Generate 10 rather than 1 feature per base class (ablation) | NIFF novel-training pipeline · MS-COCO 10-shot | AP: 29.9 -> 30.7 | Helped |
| Resample generated features during training rather than fixing the sampled features (ablation) | NIFF novel-training pipeline · MS-COCO 10-shot | AP: 25.9 -> 30.7 | Helped |
| Generate features for every base class instead of a random subset of 10 classes (ablation) | NIFF novel-training pipeline · MS-COCO 10-shot | AP: 30.5 -> 30.7 | Helped |
Closest papers by meaning
- Generalized Few-Shot Object Detection Without Forgetting (CVPR 2021 · similarity 0.89)
- General Instance Distillation for Object Detection (CVPR 2021 · similarity 0.88)
- Few-Shot Object Detection via Classification Refinement and Distractor Retreatment (CVPR 2021 · similarity 0.88)
- G-NAS: Generalizable Neural Architecture Search for Single Domain Generalization Object Detection (AAAI 2024 · similarity 0.86)
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.86)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
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.