Towards Sharper Object Boundaries in Self-Supervised Depth Estimation

2025-09-19

AI Summary

Existing self-supervised monocular depth estimation methods suffer from blurred depth discontinuities at object boundaries, leading to spurious 3D points. To address this, we propose Mixture Depth Distribution (MDD) modeling, representing per-pixel depth as a multimodal distribution—explicitly encoding uncertainty in mixture weights rather than in regression values. We further introduce a variance-aware loss and an uncertainty propagation mechanism that jointly enforce precise modeling of depth discontinuities at boundaries, all without additional annotations. Our method seamlessly integrates into standard self-supervised training pipelines. Evaluated on KITTI and VKITTIv2, MDD improves boundary sharpness by up to 35% over prior work, while significantly enhancing point cloud geometric fidelity compared to state-of-the-art methods. Notably, it is the first approach to achieve end-to-end optimizable, sharp depth discontinuity estimation without requiring fine-grained boundary annotations.

Technology Category

• Search and Optimization: Mixed Discrete/Continuous Search
• Machine Learning: Calibration & Uncertainty Quantification
• Constraint Satisfaction and Optimization: Mixed Discrete/Continuous Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Accurate monocular depth estimation is crucial for 3D scene understanding, but existing methods often blur depth at object boundaries, introducing spurious intermediate 3D points. While achieving sharp edges usually requires very fine-grained supervision, our method produces crisp depth discontinuities using only self-supervision. Specifically, we model per-pixel depth as a mixture distribution, capturing multiple plausible depths and shifting uncertainty from direct regression to the mixture weights. This formulation integrates seamlessly into existing pipelines via variance-aware loss functions and uncertainty propagation. Extensive evaluations on KITTI and VKITTIv2 show that our method achieves up to 35% higher boundary sharpness and improves point cloud quality compared to state-of-the-art baselines.

Problem

• Achieving sharp depth boundaries in monocular estimation
• Reducing blurred depth at object edges without supervision
• Modeling pixel depth as mixture distribution for uncertainty

Innovation

• Mixture distribution modeling per-pixel depth
• Variance-aware loss functions integration
• Uncertainty propagation in existing pipelines