Research map / 3D vision / Depth and stereo
GeoDepth: From Point-to-Depth to Plane-to-Depth Modeling for Self-Supervised Monocular Depth Estimation
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 GeoDepth, a self-supervised monocular depth framework that predicts per-pixel planar normals and offsets (plane-to-depth modeling) instead of depth directly, with structured plane generation and depth discontinuity awareness modules to reduce depth discontinuities.
Problem
Self-supervised monocular depth estimation treats each pixel independently (point-to-depth), causing depth discontinuities and artifacts in low-texture and planar regions.
Approach
Model the scene as a collection of planes parameterized by per-pixel normal and offset, so depth = offset / (normal^T K^-1 p); coplanar points share one representation and their depth varies only with pixel coordinates. Adds spatio-temporal normal/offset alignment losses, a plane-uniqueness gradient loss, and an adaptive photometric-threshold mask to find discontinuous regions.
Key results
- On KITTI (monocular training, 640x192), GeoDepth reaches Abs Rel 0.100 and Sq Rel 0.694, outperforming Swin-Depth (Sq Rel 0.739) and Dynamo-Depth.
- On NYUv2, GeoDepth achieves Abs Rel 0.134 and delta<1.25 of 0.833, beating StructDepth (0.142 Abs Rel).
- Ablation on KITTI shows RMSE falling from 4.471 (baseline) to 4.381 with all three components, with no extra parameters from the modules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GeoDepth instead of DIFFNet | Self-supervised monocular depth estimation · KITTI test | RMSE: 4.483 -> 4.381 | Helped |
| Use GeoDepth instead of Lite-Mono | Self-supervised monocular depth estimation trained on KITTI · Make3D test | RMSE: 6.981 -> 6.735 | Helped |
| Use GeoDepth instead of StructDepth | Self-supervised monocular depth estimation · NYUv2 test | RMSE: 0.540 -> 0.520 | Helped |
| Use GeoDepth instead of SC-Depthv1 | Self-supervised monocular depth estimation trained on NYUv2 · ScanNet test | RMSE: 0.392 -> 0.387 | Helped |
| Add Plane-to-Depth modeling (ablation) | Baseline depth model on KITTI · KITTI | RMSE: 4.471 -> 4.436 | Helped |
| Add Structured Plane Generation (ablation) | P2D model on KITTI · KITTI | RMSE: 4.436 -> 4.412 | Helped |
| Add Depth Discontinuity Awareness (ablation) | P2D+SPG model on KITTI · KITTI | RMSE: 4.412 -> 4.381 | Helped |
Limitations
- Uses a deliberately simple architecture (DepthNet and PoseNet from RA-Depth) without SOTA backbones such as MonoViT.
- Plane modeling is applied per pixel with a fixed mixing coefficient and threshold coefficient mu chosen by hand (mu=0.5).
Experimental setup
Datasets: KITTI NYUv2 Make3D ScanNet
Benchmarks: KITTI Make3D NYUv2 ScanNet
Metrics: Abs Rel Sq Rel RMSE RMSE log delta<1.25
Baselines: Monodepth2 HR-Depth CADepth-Net DIFFNet Lite-Mono Swin-Depth Dynamo-Depth StructDepth PLNet P2Net
Builds on: Monodepth2 RA-Depth Lite-Mono StructDepth
Closest papers by meaning
- SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation (AAAI 2024 · similarity 0.91)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.89)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.88)
- DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain (NeurIPS 2024 · similarity 0.88)
- SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.88)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.88)
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.