Research map / Detection and segmentation / Open-set and few-shot detection
Fetch and Forge: Efficient Dataset Condensation for 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 DCOD, the first dataset condensation framework for object detection, which trains a detector on real data (Fetch) and then synthesizes images via model inversion with foreground-background decoupling and incremental PatchExpand (Forge).
Problem
Dataset condensation has focused on image classification; extending it to object detection is hard because of multi-object localization/classification and large, high-resolution detection datasets that make bi-level optimization impractical.
Approach
Decouple bi-level dataset condensation into two stages: first store localization and classification knowledge in a trained detector's parameters (Fetch), then reconstruct synthetic images by model inversion with the detector loss, using foreground-background decoupling (random erasure on foreground, suppression on background) and incremental PatchExpand to diversify multi-instance content (Forge).
Key results
- At 1% compression on Pascal VOC, DCOD reaches 46.4% AP50 with YOLOv3-SPP (vs. 25.5% for Random).
- On MS COCO at 1% compression, DCOD achieves 12.1% mAP and 24.7% AP50, versus 8.3% mAP and 19.7% AP50 for Random.
- Ablation on VOC at 1%: adding FBD and IPE to the baseline raises AP50 from 25.9% to 46.4%; the baseline alone falls below random sampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCOD instead of Uniform sampling | YOLOv3-SPP training on Pascal VOC at 1% compression · Pascal VOC, 1% compression | AP50: 25.7±0.6 -> 46.4±0.4 | Helped |
| Use DCOD instead of Random sampling | YOLOv3-SPP training on MS COCO at 1% compression · MS COCO, 1% compression | AP50: 19.7±0.1 -> 24.7±0.2 | Helped |
| Add FBD to Forge (ablation) | DCOD on Pascal VOC at 1% compression, without IP E · Pascal VOC, 1% compression | mAP: 8.3±0.6 -> 10.0±0.3 | Helped |
| Add IP E to Forge (ablation) | DCOD on Pascal VOC at 1% compression, without FBD · Pascal VOC, 1% compression | mAP: 8.3±0.6 -> 17.7±0.4 | Helped |
| Add both FBD and IP E to Forge (ablation) | DCOD on Pascal VOC at 1% compression · Pascal VOC, 1% compression | mAP: 8.3±0.6 -> 19.8±0.4 | Helped |
| Replace random-image initialization with noise (ablation) | DCOD on Pascal VOC at 0.5% compression · Pascal VOC, 0.5% compression | mAP: 14.2±0.5 -> 6.9±0.7 | Hurt |
| Replace random-image initialization with K-center samples (ablation) | DCOD on Pascal VOC at 0.5% compression · Pascal VOC, 0.5% compression | mAP: 14.2±0.5 -> 11.9±0.6 | Hurt |
| Replace random-image initialization with Herding samples (ablation) | DCOD on Pascal VOC at 0.5% compression · Pascal VOC, 0.5% compression | mAP: 14.2±0.5 -> 11.5±0.7 | Hurt |
| Increase the αTV regularization weight from no regularization or to excessive regularization (ablation) | DCOD on Pascal VOC at 1% compression · Pascal VOC, 1% compression | performance | Mixed |
Surprising findings
- The baseline model-inversion approach without FBD and IPE underperforms random sampling.
- Initializing synthetic sets with K-center or Herding core-set selection performs worse than random initialization.
Limitations
- Not extended to more complex detectors such as DETR.
- Performance across architectures (e.g., Faster R-CNN) is still insufficient, with larger drops at low compression rates.
- Experiments are on VOC and COCO only with a single detector family evaluated for condensation.
Experimental setup
Datasets: Pascal VOC 2007+2012 trainval MS COCO
Benchmarks: Pascal VOC 2007 test MS COCO val
Metrics: mAP AP50 AP75 APs APm APl
Baselines: Random Uniform K-Center Herding
Builds on: SRe2L Dataset Distillation (DD) DeepInversion
Closest papers by meaning
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.84)
- Multi-Perspective Data Augmentation for Few-shot Object Detection (ICLR 2025 · similarity 0.84)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.84)
- M3D: Dataset Condensation by Minimizing Maximum Mean Discrepancy (AAAI 2024 · similarity 0.84)
- Elucidating the Design Space of Dataset Condensation (NeurIPS 2024 · similarity 0.84)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 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.