Research map / Label-efficient and robust learning / Self-supervised and contrastive
Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a self-supervised monocular depth method that jointly learns pre-semantic contours and depth/ego-motion, and reconstructs a variance-augmented image by concatenating a distance transform of the contours to address photometric ambiguity in low-texture regions.
Problem
Self-supervised monocular depth estimation with photometric losses is ill-posed in low-texture areas, where many correspondence candidates yield similarly low cost.
Approach
Jointly estimate pre-semantic contours (self-supervised from depth and normal Laplacian zero-crossings) and apply a distance transform to them, concatenating this map to the input image so the photometric reconstruction gains variance in texture-less areas; a theoretical argument claims the distance transform is the optimal variance-augmenting function under the constancy assumption.
Key results
- On KITTI with ResNet-50 encoder, Abs Rel 0.104 vs. 0.110 for Monodepth2 and 0.108 for FeatDepth (Shu et al.), and 0.104 with 0.725 Sq Rel.
- With DINOv2 encoder, Abs Rel 0.082, outperforming MonoViT-based variant (0.092 with MonoViT) and other listed methods.
- Contour estimation on Cityscapes reaches ODS 0.762 / OIS 0.789 / AP 0.770, above LEGO (0.710 / 0.731 / 0.729).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the distance-transform method with a DinoV2 encoder instead of RA-Depth | Monocular depth estimation · KITTI 2015 | Abs Rel: 0.096 -> 0.082 | Helped |
| Use the distance-transform method with a ResNet-50 encoder instead of MonoViT | Monocular depth estimation · KITTI 2015 | Abs Rel: 0.099 -> 0.104 | Hurt |
| Use the distance-transform method instead of CoopNet | Monocular depth estimation · Cityscapes | Abs Rel: 0.121 -> 0.115 | Helped |
| Use the distance-transform method instead of LearnK | Monocular depth estimation · Waymo / Waymo | Abs Rel: 0.157 -> 0.125 | Helped |
| Use the distance-transform method instead of CoopNet | Monocular depth estimation · Cityscapes / KITTI | Abs Rel: 0.125 -> 0.116 | Helped |
| Use the distance-transform method instead of IndoorDepth | Monocular depth estimation · NYUv2 / NYUv2 | Abs Rel: 0.126 -> 0.115 | Helped |
| Use the distance-transform method instead of MonoIndoor++ | Monocular depth estimation · NYUv2 / ScanNet | Abs Rel: 0.138 -> 0.127 | Helped |
| Add RW3 (ablation) | Depth method on KITTI 2015 · KITTI 2015 | Abs Rel: 0.113 -> 0.112 | Helped |
| Add Ldist to RW3 (ablation) | Depth method on KITTI 2015 · KITTI 2015 | Abs Rel: 0.112 -> 0.110 | Helped |
| Add Ls (ablation) | Depth method on KITTI 2015 · KITTI 2015 | Abs Rel: 0.110 -> 0.106 | Helped |
| Add edge encoder Eθ (ablation) | Depth method on KITTI 2015 · KITTI 2015 | Abs Rel: 0.106 -> 0.104 | Helped |
| Add Ls to depth and normal pseudo-label contour training (ablation) | Contour estimation · Cityscapes validation | ODS: 0.755 -> 0.760 | Helped |
Limitations
- Relies on pre-semantic contours estimated by a separate edge network, which the authors note is a dependency; future work targets variance without contours.
- Theoretical optimality result holds under the Eikonal constraint and simplified convex-shape assumptions.
- Comparisons with some methods use off-the-shelf semantic algorithms or different encoders, complicating fair comparison.
Experimental setup
Datasets: KITTI Cityscapes Waymo NYUv2 KITTI MOTS
Benchmarks: KITTI Cityscapes Waymo NYUv2 ScanNet KITTI MOTS
Metrics: Abs Rel Sq Rel RMSE RMSE log delta accuracy ODS OIS AP
Baselines: Monodepth2 CoopNet Guizilini et al. Shu et al. (FeatDepth) MonoViT HR-Depth RA-Depth DIFFNet LEGO LearnK Struct2Depth IndoorDepth
Builds on: Monodepth2 CoopNet LEGO MonoViT Struct2Depth
Closest papers by meaning
- Semantically-Guided Representation Learning for Self-Supervised Monocular Depth (ICLR 2020 · similarity 0.89)
- Mining Supervision for Dynamic Regions in Self-Supervised Monocular Depth Estimation (CVPR 2024 · similarity 0.88)
- From-Ground-To-Objects: Coarse-to-Fine Self-supervised Monocular Depth Estimation of Dynamic Objects with Ground Contact Prior (CVPR 2024 · similarity 0.88)
- Mind The Edge: Refining Depth Edges in Sparsely-Supervised Monocular Depth Estimation (CVPR 2024 · similarity 0.87)
- SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation (AAAI 2024 · similarity 0.87)
- Intrinsic Image Decomposition for Robust Self-supervised Monocular Depth Estimation on Reflective Surfaces (AAAI 2025 · similarity 0.87)
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.