Research map / Detection and segmentation / Open-set and few-shot detection
ReCon: Region-Controllable Data Augmentation with Rectification and Alignment 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 ReCon, a training-free augmentation framework that adds region-guided rectification (grounding-model feedback injected into diffusion sampling) and region-aligned cross-attention to structure-controllable diffusion models to produce better-aligned object detection training data.
Problem
Generative data augmentation for object detection suffers from misgenerated regions, content-position mismatch, and semantic leakage, and existing layout-controlled methods often need costly fine-tuning on large datasets.
Approach
During diffusion sampling, a frozen perception model (Grounded-SAM) detects misgenerated instances via IoU matching against annotations at several timesteps, and those regions are replaced with noised original latents (rectification); cross-attention is restricted so each region attends only to its own class prompt (region-aligned cross-attention). No generator retraining is needed.
Key results
- ControlNet + ReCon reaches 35.5 mAP on COCO versus 34.9 for ControlNet and 34.8 for GeoDiffusion, and GLIGEN improves from 34.6 to 35.3 with ReCon.
- In 10% COCO data-scarce setting, ReCon raises mAP from 18.5 (real only) to 21.7 and with RandAugment to 22.0.
- On DEIM-D-FINE-N, ReCon improves mAP from 38.5 to 39.8 on COCO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ReCon to ControlNet | ControlNet with Faster R-CNN R-50-FPN · COCO | mAP: 34.9 -> 35.5 | Helped |
| Add ReCon to GLIGEN | GLIGEN with Faster R-CNN R-50-FPN · COCO | mAP: 34.6 -> 35.3 | Helped |
| Add ReCon to Instance Diffusion | Instance Diffusion with Faster R-CNN R-50-FPN · COCO | mAP: 35.0 -> 35.6 | Helped |
| Augment the 1% training subset with ReCon | Faster R-CNN R-50-FPN · COCO 1% data | mAP: 2.5 -> 3.9 | Helped |
| Augment the 5% training subset with ReCon | Faster R-CNN R-50-FPN · COCO 5% data | mAP: 15.9 -> 16.7 | Helped |
| Augment the 10% training subset with ReCon | Faster R-CNN R-50-FPN · COCO 10% data | mAP: 21.2 -> 21.7 | Helped |
| Augment VOC training data with ReCon | Faster R-CNN · PASCAL VOC 2007 test | mAP: 77.8 -> 78.5 | Helped |
| Add ReCon to the 30-shot training data at 2× expansion | YOLOX-S · COCO, 30-shot | mAP: 5.4 -> 6.7 | Helped |
| Add region-guided rectification (RGR) (ablation) | ControlNet with Faster R-CNN R-50-FPN · COCO | mAP: 34.9 -> 35.3 | Helped |
| Add region-aligned cross-attention (RACA) to RGR (ablation) | ControlNet with Faster R-CNN R-50-FPN · COCO | mAP: 35.3 -> 35.5 | Helped |
| Use x_{0|(t-N)} instead of x_t as the perception target (ablation) | ReCon · COCO | mAP: 35.0 -> 35.5 | Helped |
| Use the four-stage rectification schedule [0.75, 0.5, 0.25, 0.1]T rather than rectification only at 0.75T (ablation) | ReCon · COCO | mAP: 35.2 -> 35.5 | Helped |
Surprising findings
- Early-stage-only rectification gives modest gains, and performance saturates beyond four rectification stages.
- Masked attention alone still suffers semantic leakage because text features of different categories are not disentangled.
Limitations
- Adds computation time that grows with data volume and requires an extra perception model.
- Quality depends on the strength of the perception model used for rectification.
- Results are reported as improvements over ControlNet-based baselines, with gains on COCO of roughly 0.6 mAP.
Experimental setup
Datasets: COCO PASCAL VOC 2007 PASCAL VOC 2012
Benchmarks: COCO PASCAL VOC 2007 test
Metrics: mAP AP50 AP75 APm APl mAR FID
Baselines: ControlNet GLIGEN GeoDiffusion DetDiffusion Instance Diffusion Layout Diffusion ReCo RandAugment SDEdit
Builds on: ControlNet Stable Diffusion Grounded-SAM InstanceDiffusion cache-based fast sampling
Closest papers by meaning
- Reconstruction Alignment Improves Unified Multimodal Models (ICLR 2026 · similarity 0.87)
- Generative Region-Language Pretraining for Open-Ended Object Detection (CVPR 2024 · similarity 0.86)
- Plug-and-Play Context Feature Reuse for Efficient Masked Generation (NeurIPS 2025 · similarity 0.86)
- Decompositional Neural Scene Reconstruction with Generative Diffusion Prior (CVPR 2025 · similarity 0.85)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.85)
- RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation (NeurIPS 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.