Research map / Label-efficient and robust learning / Recognition and long-tail
CNN-Based Patch Matching for Optical Flow With Thresholded Hinge Embedding Loss
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 thresholded hinge loss with t = 0.3 instead of ordinary hinge embedding loss | Siamese CNN patch matching · KITTI 2012 validation | EPE all: 3.33 px -> 2.60 px | Helped |
| Use thresholded hinge loss with t = 0.3 instead of gap loss | Siamese CNN patch matching · KITTI 2012 validation | EPE all: 3.36 px -> 2.60 px | Helped |
| Reject zero-loss samples without backpropagation | Thresholded hinge training · Training | training speed: factor of around two | Helped |
| Add a 2x low-pass filter to feature maps (ablation) | CNN multi-scale patch matching · KITTI 2012 validation | EPE all: 2.80 px -> 2.60 px | Helped |
| Use 1x, 1x, 2x, 4x scales rather than 1x, 2x, 4x, 8x (ablation) | CNN multi-scale feature creation · KITTI 2012 validation | EPE all: 2.98 px -> 2.60 px | Helped |
| Train the multi-resolution CNN on 100%, 50%, and 25% resolutions (ablation) | CNN multi-scale feature creation · KITTI 2012 validation | EPE all: 2.80 px -> 2.60 px | Helped |
| Train the multi-resolution CNN on 100%, 50%, and 25% resolutions rather than only 50% and 25% (ablation) | CNN multi-scale feature creation · KITTI 2012 validation | EPE all: 2.84 px -> 2.60 px | Helped |
| Use threshold t = 0.2 rather than t = 0.3 (ablation) | Thresholded hinge loss · KITTI 2012 validation | EPE all: 2.60 px -> 3.10 px | Hurt |
| Use threshold t = 0.4 rather than t = 0.3 (ablation) | Thresholded hinge loss · KITTI 2012 validation | EPE all: 2.60 px -> 3.14 px | Hurt |
| Use threshold t = 0.45 rather than t = 0.3 (ablation) | Thresholded hinge loss · KITTI 2012 patch matching robustness | relative matching robustness error | Hurt |
| Use DrLIM instead of thresholded hinge loss | Siamese CNN patch matching · KITTI 2012 validation | EPE all: 2.60 px -> 2.79 px | Hurt |
| Use Hard Mining x2 instead of thresholded hinge loss | Siamese CNN patch matching · KITTI 2012 validation | EPE all: 2.60 px -> 2.99 px | Hurt |
Closest papers by meaning
- PatchBatch: A Batch Augmented Loss for Optical Flow (CVPR 2016 · similarity 0.93)
- Optical Flow Requires Multiple Strategies (but Only One Network) (CVPR 2017 · similarity 0.88)
- A Theory on Flow Matching with Neural Networks (NeurIPS 2026 · similarity 0.86)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.85)
- Context-Aware Iteration Policy Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.85)
- Neuroexplicit Diffusion Models for Inpainting of Optical Flow Fields (ICML 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.