Research map / Detection and segmentation / CNN features
Delving into Cascaded Instability: A Lipschitz Continuity View on Image Restoration and Object Detection Synergy
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.
Analyzes Lipschitz continuity mismatch between image restoration and object detection networks, and proposes LROD/LR-YOLO, which shares YOLO backbone features with a restoration head and adds a parameter-space gradient-norm regularizer to improve detection under haze and low light.
Problem
Restoration-then-detection cascades for adverse conditions are unstable because detectors have much higher Lipschitz constants than restoration networks, so restoration noise is amplified during detection and optimization is disrupted.
Approach
Integrate restoration learning into the detector's backbone (sharing the first three stages with a restoration head) so the detector is regularized toward the smoother restoration mapping, plus a parameter-space gradient-norm penalty, instead of a sequential cascade.
Key results
- LR-YOLOv8 reaches 53.2 mAP on RTTS vs 45.3 for YOLOv8 and 50.3 for the best cascade (GDIP→YOLOv8).
- LR-YOLOv8 reaches 71.7 mAP on VOC_Dark_Val vs 63.4 for YOLOv8 and 68.9 for the best prior method (GDIP→YOLOv8).
- Ablation: adding restoration learning and parameter-space smoothing gives +3.9 mAP on RTTS for YOLOv8 relative to baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate restoration learning and parameter-space smoothing into YOLOv10 (ablation) | YOLOv10 · RTTS | mAP: +3.2 | Helped |
| integrate restoration learning and parameter-space smoothing into YOLOv8 (ablation) | YOLOv8 · RTTS | mAP: +3.9 | Helped |
| integrate restoration learning and parameter-space smoothing into YOLOv10 (ablation) | YOLOv10 · ExDark | mAP: +3.2 | Helped |
| integrate restoration learning and parameter-space smoothing into YOLOv8 (ablation) | YOLOv8 · ExDark | mAP: +2.9 | Helped |
| share only stages F1–F2 instead of F1–F3 (ablation) | LR-YOLOv8 · RTTS | mAP: 53.2 -> 51.9 | Hurt |
| share stages F1–F4 instead of F1–F3 (ablation) | LR-YOLOv8 · RTTS | mAP: 53.2 -> 52.8 | Hurt |
| replace the proposed regularization with spectral norm regularization (ablation) | YOLOv8 · RTTS | mAP: 53.2 -> 50.1 | Hurt |
| replace the proposed regularization with PGD adversarial training (ablation) | YOLOv8 · RTTS | mAP: 53.2 -> 40.8 | Hurt |
| vary regularization coefficients across the reported sweep (ablation) | LR-YOLOv8 · RTTS | mAP: 52.8–53.2 | No clear effect |
| integrate LR-YOLO into detection pipelines | YOLOv10 · RTTS | mAP: 47.2 -> 49.2 | Helped |
| integrate LR-YOLO into detection pipelines | YOLOv8 · RTTS | mAP: 50.3 -> 53.2 | Helped |
| integrate LR-YOLO into detection pipelines | YOLOv8 · COCO_Haze_Val | mAP: +1.0 | Helped |
Surprising findings
- Sharing the backbone up to stage F3 works better than sharing deeper (F1–F4), which the authors attribute to task interference.
Limitations
- Assumes each input contains only a single type of degradation.
- Training relies on synthetically degraded VOC/COCO images; real-world results are out-of-domain only.
Experimental setup
Datasets: Pascal VOC COCO RTTS ExDark
Benchmarks: VOC_Haze_Val RTTS VOC_Dark_Val ExDark COCO_Haze_Val COCO_Dark_Val
Metrics: mAP mAP_difficult AP50 AP75 APS APM APL PSNR LPIPS Params
Baselines: YOLOv10 YOLOv8 SFNet ConvIR LLFormer Retinexformer IA-YOLO GDIP FeatEnHancer InstructIR RT-DETR Faster R-CNN
Builds on: YOLOv8 YOLOv10 ConvIR Charbonnier loss
Closest papers by meaning
- Focus, Align, and Sustain: Counteracting Gradient Dilution in Incremental Object Detection (ICML 2026 · similarity 0.85)
- Consistency Beyond Contrast: Enhancing Open-Vocabulary Object Detection Robustness via Contextual Consistency Learning (CVPR 2026 · similarity 0.85)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.84)
- YOLOv10: Real-Time End-to-End Object Detection (NeurIPS 2024 · similarity 0.84)
- DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.84)
- Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond (AAAI 2024 · similarity 0.83)
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.