Research map / Detection and segmentation / Open-set and few-shot detection
Improved Handling of Motion Blur in Online Object Detection
Mohamed Sayed, Gabriel Brostow
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| expand bounding-box labels to cover the blur-kernel displacement (ablation) | blur-augmented Faster R-CNN · COCO minival E5, P1 blur | mAP@0.5: 31.37 -> 38.35 | Improved |
| specialize detector networks by exposure and select a network using a blur estimator (ablation) | expanded-label blur-augmented Faster R-CNN · COCO minival E5, P1 blur | mAP@0.5: 38.35 -> 43.78 | Improved |
| specialize networks by blur type instead of exposure (ablation) | expanded-label blur-augmented Faster R-CNN · COCO minival E5, P1 blur | mAP@0.5: 43.78 -> 40.81 | Hurt |
| center blur kernels to align blurred objects with detection labels (ablation) | blur-augmented detector · COCO minival, motion blur | mAP@0.5: up to 8-10mAP@50 | Improved |
| deblur the image before detection (ablation) | Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 4.58 -> 6.26 | Improved |
| undersample the input along blur axes and reverse-scale backbone activations (ablation) | Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 4.58 -> 31.39 | Improved |
| apply spatial AugMix without adapting the labels (ablation) | blur-augmented Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 4.58 -> 8.21 | Improved |
| apply pixel-level AugMix (ablation) | Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 4.58 -> 4.48 | No clear effect |
| adapt batch-normalization statistics from the test minibatch (ablation) | Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 4.58 -> 4.40 | Hurt |
| train with uncentered blur kernels (ablation) | blur-augmented Faster R-CNN · COCO minival E5, P1 blur, standard labels | mAP@0.5: 31.37 -> 25.66 | Hurt |
| combine blur augmentation with non-spatial AugMix (ablation) | blur-augmented detector · RealBlur blurry | mAP@0.5: 35.63 -> 36.82 | Improved |
| apply minibatch normalization to the expanded-label model (ablation) | expanded-label blur-augmented Faster R-CNN · COCO minival E5, P1 blur, expanded labels | mAP@0.5: 38.35 -> 13.88 | Hurt |
Closest papers by meaning
- Improving Robustness of Semantic Segmentation to Motion-Blur Using Class-Centric Augmentation (CVPR 2023 · similarity 0.86)
- Retrieval Robust to Object Motion Blur (ECCV 2024 · similarity 0.86)
- Objects Do Not Disappear: Video Object Detection by Single-Frame Object Location Anticipation (ICCV 2023 · similarity 0.85)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.85)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · 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.
