Research map / Architectures and efficiency / RNNs and normalization
PatchBatch: A Batch Augmented Loss for Optical Flow
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use the CENTRIFUGE+SD loss instead of SPRING (ablation) | PatchBatch · KITTI2012 validation | error > 3 pixels after ANN: 10.04 -> 8.54 | Helped |
| Use the CENTRIFUGE+SD loss instead of SPRING (ablation) | PatchBatch · KITTI2012 validation | error > 3 pixels after interpolation: 4.96 -> 4.97 | No clear effect |
| Use the CENTRIFUGE+SD loss instead of SPRING (ablation) | PatchBatch · KITTI2015 validation | error > 3 pixels after ANN: 28.74 -> 26.34 | Helped |
| Use the CENTRIFUGE+SD loss instead of SPRING (ablation) | PatchBatch · MPI-Sintel FINAL validation | EPE all: 3.06 -> 3.15 | Mixed |
| Add fine-grained batch normalization instead of conventional batch normalization (ablation) | PatchBatch · KITTI2012 validation | error > 3 pixels after interpolation: 5.3 -> 4.80 | Helped |
| Increase patch size to 71×71 (ablation) | PatchBatch ACCURATE · KITTI2012 validation | error > 3 pixels after interpolation: 4.80 -> 4.79 | Helped |
| Reduce descriptor size to 32D (ablation) | PatchBatch ACCURATE · KITTI2012 validation | error > 3 pixels after interpolation: 4.80 -> 5.23 | Hurt |
| Increase descriptor size to 1024D (ablation) | PatchBatch ACCURATE · KITTI2012 validation | error > 3 pixels after interpolation: 4.80 -> 4.81 | No clear effect |
| Train on KITTI2015 rather than KITTI2012 (ablation) | PatchBatch ACCURATE · KITTI2012 validation | error > 3 pixels after interpolation: 4.80 -> 4.97 | Hurt |
| Add the batch standard-deviation term to the loss (ablation) | DrLIM metric learning · Synthetic Gaussian experiments | AUC | Helped |
| Replace the spring DrLIM loss with the centrifuge loss (ablation) | DrLIM metric learning · Synthetic Gaussian experiments with fixed-norm samples | AUC | Helped |
| Replace the spring DrLIM loss with the centrifuge loss (ablation) | DrLIM metric learning · Synthetic Gaussian experiments without fixed-norm samples | AUC | Hurt |
Closest papers by meaning
- CNN-Based Patch Matching for Optical Flow With Thresholded Hinge Embedding Loss (CVPR 2017 · similarity 0.93)
- Optical Flow Requires Multiple Strategies (but Only One Network) (CVPR 2017 · similarity 0.88)
- DIP: Deep Inverse Patchmatch for High-Resolution Optical Flow (CVPR 2022 · similarity 0.88)
- Learning Optical Flow With Kernel Patch Attention (CVPR 2022 · similarity 0.86)
- Iterative Residual Refinement for Joint Optical Flow and Occlusion Estimation (CVPR 2019 · similarity 0.86)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 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.