Research map / 3D vision / Depth and stereo
Mind The Edge: Refining Depth Edges in Sparsely-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 an edge loss that supervises MDE depth-edge localization using edges predicted by a detector trained on synthetic dense-depth data, and releases manually annotated depth-edge sets for KITTI and DDAD.
Problem
LIDAR-supervised monocular depth estimation produces inaccurate depth edges (thick or misplaced discontinuities) because sparse LIDAR GT gives no direct supervision for edge locations, hurting applications like novel view synthesis and AR occlusion.
Approach
Train a Depth Edge Estimation network on dense synthetic depth (GTA-PreSIL) with Canny-derived edge GT, infer approximate depth-edge maps on real training data, and use a differentiable edge detection block with a balanced BCE edge loss to sharpen predicted depth edges while keeping per-pixel accuracy.
Key results
- On KITTI-DE, Packnet-SAN + EL raises edge AUC from 47.56% to 61.87% while ARE is 3.61% vs 3.45% for the baseline.
- On DDAD-DE, edge AUC rises from 31.52% (Packnet-SAN) to 48.32% with ARE 8.99% vs 8.89%.
- On the standard KITTI test set, PixelFormer + EL achieves ARE 5.59% vs 5.46% baseline, indicating comparable per-pixel accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Edge Loss | Packnet-SAN · KITTI-DE | AUC (edges): 47.56% -> 61.87% | Helped |
| add Edge Loss | AdaBins · KITTI-DE | AUC (edges): 41.23% -> 53.47% | Helped |
| add Edge Loss | PixelFormer · KITTI-DE | AUC (edges): 32.79% -> 46.23% | Helped |
| add Edge Loss | Packnet-SAN · KITTI-DE | ARE: 3.45% -> 3.61% | Hurt |
| add Edge Loss | AdaBins · KITTI-DE | ARE: 3.14% -> 3.11% | Helped |
| add Edge Loss | PixelFormer · KITTI-DE | ARE: 3.00% -> 2.94% | Helped |
| add Edge Loss | Packnet-SAN · KITTI test | ARE: 6.17% -> 6.50% | Hurt |
| add Edge Loss | AdaBins · KITTI test | ARE: 6.28% -> 6.21% | Helped |
| add Edge Loss | PixelFormer · KITTI test | ARE: 5.45% -> 5.59% | Hurt |
| add Edge Loss | Packnet-SAN · DDAD-DE | AUC (edges): 31.52% -> 48.32% | Helped |
| add Edge Loss | Packnet-SAN · DDAD-DE | ARE: 8.89% -> 8.99% | Hurt |
| add Edge Loss | Packnet-SAN · DDAD test | ARE: 9.49% -> 10.0% | Hurt |
Surprising findings
- Edge quality is not strongly correlated with per-pixel depth quality, and on DDAD the per-pixel ARE slightly worsens (9.49% to 10.0% on the standard test set).
- BoostingDepth and GradientFusion merging methods perform slightly worse or significantly worse than the Packnet-SAN baseline on edge AUC on KITTI.
Limitations
- Slight per-pixel depth accuracy decrease on DDAD test (ARE 9.49% to 10.0%) and KITTI test for some methods.
- Domain gap between synthetic GTA-PreSIL training of the DEE network and real target data; the DEE model is not adapted to the target domain.
- Depth edge evaluation sets are small (102 KITTI and 50 DDAD images) and manually annotated.
Experimental setup
Datasets: GTA-PreSIL KITTI DDAD Synscapes HR-WSI
Benchmarks: KITTI-DE KITTI test (Eigen split) DDAD-DE DDAD test Synscapes
Metrics: AUC (edges) ORD ARE delta < 1.25 precision-recall
Baselines: Packnet-SAN AdaBins PixelFormer BoostingDepth GradientFusion
Builds on: Packnet-SAN AdaBins PixelFormer Canny edge detector
Closest papers by meaning
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.90)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.89)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.88)
- Flexible Depth Completion for Sparse and Varying Point Densities (CVPR 2024 · similarity 0.88)
- DME: Unveiling the Bias for Better Generalized Monocular Depth Estimation (AAAI 2024 · similarity 0.87)
- Improved Monocular Depth Prediction Using Distance Transform Over Pre-semantic Contours with Self-supervised Neural Networks (CVPR 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.