Research map / 3D vision / Depth and stereo
Semi-Supervised Adversarial Monocular Depth Estimation
Rongrong Ji, Ke Li, Yan Wang, Xiaoshuai Sun, Feng Guo, Xiaowei Guo, Yongjian Wu, Feiyue Huang, Jiebo Luo
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 |
|---|---|---|---|
| add pair discriminator and depth discriminator together (vs depth discriminator only) (ablation) | semi-supervised GAN loss · NYU Depth v2 (500 GT) | rel: 0.191 -> 0.183 | Improved |
| add semi-supervised GAN loss vs L1 supervised loss | Hu et al. generator · NYU Depth v2 (500 GT) | rel: 0.201 -> 0.183 | Improved |
| add semi-supervised framework to Hu et al. | Hu et al. · NYU Depth v2 (100 GT) | rel: 0.322 -> 0.232 | Improved |
| add semi-supervised framework to FCRN | FCRN · NYU Depth v2 (500 GT) | rel: 0.272 -> 0.254 | Improved |
| add semi-supervised framework to Unet | Unet · NYU Depth v2 (1000 GT) | rel: 0.365 -> 0.369 | No clear effect |
| add unlabeled RGB images beyond 10x the labeled pairs (ablation) | semi-supervised model · NYU Depth v2 (500 GT) | depth accuracy: gain saturates after 10 times more RGB images | No clear effect |
| add unlabeled RGB images within 1/2k to 10k range (ablation) | semi-supervised model · NYU Depth v2 (500 GT) | depth accuracy | Improved |
| use semi-supervised framework with abundant labeled data (>3.5K pairs) | semi-supervised model · NYU Depth v2 | depth accuracy | No clear effect |
| add 425 unlabeled RGB images in semi-supervised training | Ours-GAN · Make3D | rel: 0.158 -> 0.153 | Improved |
| add 20k Cityscape RGB images | Ours-GAN · KITTI Eigen split | rel: 0.107 -> 0.093 | Improved |
| add 400 unlabeled Make3D RGB images in training (domain adaptation) | Ours · KITTI to Make3D | rel: 0.555 -> 0.447 | Improved |
| use GAN loss with sufficient data (12K) vs Zhang et al. | Ours · NYU Depth v2 (12K GT) | delta1: 0.824 -> 0.851 | Mixed |
Closest papers by meaning
- Unsupervised Monocular Depth Estimation With Left-Right Consistency (CVPR 2017 · similarity 0.91)
- BRNet: Exploring Comprehensive Features for Monocular Depth Estimation (ECCV 2022 · similarity 0.90)
- Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue (ECCV 2016 · similarity 0.90)
- AdaDepth: Unsupervised Content Congruent Adaptation for Depth Estimation (CVPR 2018 · similarity 0.90)
- Semi-Supervised Deep Learning for Monocular Depth Map Prediction (CVPR 2017 · similarity 0.90)
- Improving Domain Generalization in Self-Supervised Monocular Depth Estimation via Stabilized Adversarial Training (ECCV 2024 · similarity 0.90)
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.
