Research map / 3D vision / Depth and stereo
Multi-Scale Continuous CRFs as Sequential Deep Networks for Monocular Depth Estimation
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 |
|---|---|---|---|
| Fuse CNN side outputs with a 5-scale multi-scale CRF instead of traditional side-output fusion | VGG-CD depth prediction · NYU Depth V2 | rel: 0.185 -> 0.163 | Helped |
| Use multi-scale CRF rather than cascade CRF | 5-scale VGG-CD depth prediction · NYU Depth V2 | rel: 0.169 -> 0.163 | Helped |
| Increase the number of scales from 3 to 5 in the multi-scale CRF (ablation) | VGG-CD depth prediction · NYU Depth V2 | rel: 0.172 -> 0.163 | Helped |
| Increase the number of scales from 3 to 5 in the cascade CRF (ablation) | VGG-CD depth prediction · NYU Depth V2 | rel: 0.176 -> 0.169 | Helped |
| Use inner-to-outer rather than outer-to-inner cascade order (ablation) | Cascade depth refinement · NYU Depth V2 | rel: 0.175 -> 0.169 | Helped |
| Add the CRF fusion module to the pretrained front-end network | ResNet50 · NYU Depth V2 | rel: 0.168 -> 0.143 | Helped |
| Pretrain the front end with 95K rather than 4.7K image pairs (ablation) | ResNet50 cascade model · NYU Depth V2 | rel: 0.143 -> 0.121 | Helped |
| Use ResNet50 rather than VGG-CD as the CRF front-end architecture (ablation) | CRF depth prediction · NYU Depth V2 | rel: 0.169 -> 0.143 | Helped |
| Increase mean-field iterations beyond 5 (ablation) | Both CRF models · NYU Depth V2 | depth prediction performance | No clear effect |
| Use the multi-scale CRF rather than cascade CRF | ResNet50 depth prediction · Make3D C1 | rel: 0.213 -> 0.206 | Helped |
| Increase pretraining data to approximately 10K image pairs for the multi-scale CRF (ablation) | ResNet50 depth prediction · Make3D C1 | rel: 0.206 -> 0.184 | Helped |
Closest papers by meaning
- Structured Attention Guided Convolutional Neural Fields for Monocular Depth Estimation (CVPR 2018 · similarity 0.92)
- End-To-End Training of Hybrid CNN-CRF Models for Stereo (CVPR 2017 · similarity 0.88)
- Neural Window Fully-Connected CRFs for Monocular Depth Estimation (CVPR 2022 · similarity 0.88)
- Learning Deep Structured Multi-Scale Features using Attention-Gated CRFs for Contour Prediction (NeurIPS 2017 · similarity 0.87)
- Boosting the Performance of Generic Deep Neural Network Frameworks with Log-supermodular CRFs (NeurIPS 2022 · similarity 0.86)
- Gaussian Conditional Random Field Network for Semantic Segmentation (CVPR 2016 · similarity 0.86)
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.