Research map / Detection and segmentation / 3D detection and driving
SGM-Nets: Semi-Global Matching With Neural Networks
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 |
|---|---|---|---|
| Replace hand-tuned penalties with learned signed SGM-Net penalties | SGM with MC-CNN matching costs · SceneFlow Driving test (scene backwards) | Out-Noc error: 22.0 -> 18.3 | Helped |
| Use signed rather than standard penalty parameterization (ablation) | SGM-Net with MC-CNN matching costs · SceneFlow Driving test (scene backwards) | Out-Noc error: 19.5 -> 18.3 | Helped |
| Add neighbor cost to the path-cost training objective (ablation) | Signed SGM-Net · SceneFlow Driving test (scene backwards), MC-CNN | Out-Noc error: 18.3 -> 18.3 | No clear effect |
| Use path cost rather than neighbor cost alone (ablation) | Signed SGM-Net · SceneFlow Driving test (scene backwards), MC-CNN | Out-Noc error: 22.5 -> 18.3 | Helped |
| Add patch-position input (ablation) | Standard SGM-Net · SceneFlow Driving test (scene backwards), MC-CNN | Out-Noc error: 20.1 -> 19.5 | Helped |
| Train the network rather than use initial penalties (ablation) | Signed SGM-Net · SceneFlow Driving test (scene backwards), MC-CNN | Out-Noc error: 32.9 -> 18.3 | Helped |
| Include real KITTI 2012 training images in addition to synthetic training data (ablation) | SGM-Net · KITTI 2015 training images | D1-all: 5.09% -> 4.49% (Standard); 5.58% -> 4.38% (Signed) | Helped |
| Use signed SGM-Net rather than standard SGM-Net (ablation) | SGM on KITTI 2012 training images · KITTI 2012 training images | Out-Noc error (SGM): 3.10% -> 2.99% | Helped |
| Use signed SGM-Net rather than standard SGM-Net (ablation) | SGM on KITTI 2015 training images · KITTI 2015 training images | D1-all (SGM): 4.49% -> 4.38% | Helped |
| Use SGM-Net penalties instead of hand-tuned dynamic penalties | SGM on KITTI training/evaluation splits · KITTI 2012 and KITTI 2015 training images | Out-Noc error / D1-all: 3.19% -> 2.99%; 4.72% -> 4.38%; 3.23% -> 3.00%; 5.18% -> 4.76% | Mixed |
| Use signed SGM-Net rather than standard SGM-Net on the KITTI 2015 test set (ablation) | SGM · KITTI 2015 test | D1-all: 3.09% -> 3.18% | Hurt |
Closest papers by meaning
- SMD-Nets: Stereo Mixture Density Networks (CVPR 2021 · similarity 0.86)
- GIM: Learning Generalizable Image Matcher From Internet Videos (ICLR 2024 · similarity 0.86)
- GA-Net: Guided Aggregation Net for End-To-End Stereo Matching (CVPR 2019 · similarity 0.86)
- Semi-Supervised Deep Learning for Monocular Depth Map Prediction (CVPR 2017 · similarity 0.86)
- Guided Stereo Matching (CVPR 2019 · similarity 0.85)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 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.