Generalizable Depth Perception with Everyday Sensors

Research Overview:

This research series aims to build a unified and generalizable depth perception framework that works reliably across diverse environments and sensors — from LiDAR and ToF to stereo and monocular cameras.  The core goal is to bridge the gap between sparse, sensor-dependent depth data and dense, universal depth understanding through foundation model integration and geometry-aware representations.

Key Contributions Across Publications:

1. Zero-shot Depth Completion (ICCV 2025 Highlight, 2.3% Total)

   • Introduces Test-Time Prompt Tuning, a parameter-efficient method that adapts depth foundation models using sparse sensor measurements only.

   • Eliminates the need for dense ground truth or retraining, achieving real-time zero-shot adaptation across unseen environments.

2. Universal Depth Completion (NeurIPS 2024)

   • Defines a universal depth completion problem acknowledging domain diversity between indoor and outdoor environments.

   • Introduces a lightweight baseline that combines RGB cues and sparse sensor data for rapid adaptation.

   • Builds hierarchical 3D structures in hyperbolic space to improve generalization under few-shot or unseen conditions.
     
     

3. Sensor-agnostic Depth Estimation (CVPR 2024)

   • Proposes Depth Prompting to enable monocular foundation models to understand sensor-specific depth distributions.

   • Allows pretrained models to output absolute metric-scale depth without retraining, overcoming sensor range constraints.
     
     

4. Hyperbolic Spatial Affinity (ICML 2023)

   • Models pixel-wise geometric relationships in hyperbolic space, capturing hierarchical structures in depth propagation.

   • Enhances robustness in regions with discontinuity and sparse supervision.