Research map / Detection and segmentation / Open-set and few-shot detection
Improving Robustness of Semantic Segmentation to Motion-Blur Using Class-Centric Augmentation
Aakanksha, A. N. Rajagopalan
Read the paperCheck 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 |
|---|---|---|---|
| Apply class-centric motion-blur augmentation instead of the strongest baseline | DeepLabv3+ ResNet-50 · PASCAL VOC L1, L2, L3 | mIoU: 2.3%, 3.5% and 3.2% | Improved |
| Apply class-centric motion-blur augmentation instead of the strongest baseline | DeepLabv3+ ResNet-50 · Cityscapes L1, L2, L3 | mIoU: 1.9%, 3.4% and 3.1% | Improved |
| Apply CCMBA rather than the best baseline at each blur level | Segformer MiT-B2 · Cityscapes L1, L2, L3 | mIoU | No clear effect |
| Apply CCMBA rather than PbN [14] | DeepLabv3+ · Cityscapes-C S1, S2, S3 | mIoU: 2.3%, 5.2% and 9% | Improved |
| Apply CCMBA rather than pre-trained Segformer | Segformer · Cityscapes-C S2, S3 | mIoU: 1.7%, 5.1 % | Improved |
| Apply CCMBA rather than MBA | Segformer · Cityscapes-C clean and S3 | mIoU: 2%, 0.5% | Improved |
| Initialize CCMBA training from clean pre-trained weights rather than random initialization (ablation) | DeepLabv3+ · PASCAL VOC clean, L1, L2, L3 | mIoU: Random: 75.5, 73.6, 69.6, 61.5; Pre-trained: 76.5, 74.6, 72.1, 66.0 | Improved |
| Use a larger backbone (ablation) | CCMBA with DeepLabv3+ · PASCAL VOC clean and L1, L2, L3 | mIoU: MobileNetV2: 69.3, 68.2, 65.6, 61.5; ResNet-50: 76.3, 75.3, 72.5, 68.6; ResNet-101: 77.6, 76.3, 75.2, 71.3 | Improved |
| Train on all three blur levels rather than one blur level (ablation) | CCMBA with DeepLabv3+ MobileNetV2 · PASCAL VOC clean and L1, L2, L3 | mIoU: Clean: 69.3; L1: 68.2; L2: 66.6; L3: 61.5 | Improved |
| Use only clean training images rather than blurred augmentation (ablation) | DeepLabv3+ MobileNetV2 · PASCAL VOC L1, L2, L3 | mIoU: 60.3, 44.4, 29.0 | Hurt |
| Train using only L1 blur rather than all three blur levels (ablation) | CCMBA with DeepLabv3+ MobileNetV2 · PASCAL VOC L1, L2, L3 | mIoU: 65.1, 58.7, 39.1 | Hurt |
| Train using only L2 blur rather than all three blur levels (ablation) | CCMBA with DeepLabv3+ MobileNetV2 · PASCAL VOC L1, L2, L3 | mIoU: 65.6, 63.2, 53.5 | Hurt |
Closest papers by meaning
- From Motion Blur to Motion Flow: A Deep Learning Solution for Removing Heterogeneous Motion Blur (CVPR 2017 · similarity 0.86)
- Video Deblurring via Semantic Segmentation and Pixel-Wise Non-Linear Kernel (ICCV 2017 · similarity 0.86)
- Improved Handling of Motion Blur in Online Object Detection (CVPR 2021 · similarity 0.86)
- Modeling the Background for Incremental and Weakly-Supervised Semantic Segmentation (TPAMI 2021 · similarity 0.86)
- Unlocking the Potential of Pre-trained Vision Transformers for Few-Shot Semantic Segmentation through Relationship Descriptors (CVPR 2024 · similarity 0.85)
- A Surprisingly Simple Approach to Generalized Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.85)
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.
