Self-Distilled Depth Refinement with Noisy Poisson Fusion

Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao§, Jianming Zhang

Paper Github Poster

• We model the depth refinement task through the noisy Poisson fusion problem with local inconsistency noise and edge deformation noise as two types of depth degradation.
• We present the robust and efficient Self-distilled Depth Refinement (SDDR) framework, which can generate accurate depth edge representation by the coarse-to-fine self-distillation paradigm.
• We design the edge-guided gradient loss and edge-based fusion loss, as the edge-based guidance to enforce the model with both consistent depth structures and meticulous depth edges.

Towards Robust Monocular Depth Estimation in Non-Lambertian Surfaces

Junrui Zhang, Jiaqi Li§, Yachuan Huang, Yiran Wang, Jinghong Zheng, Liao Shen, Zhiguo Cao

Paper Github

We focus on the robustness of depth estimation on mirrors and transparent objects, and based on the existing monocular depth estimation model, we complement its robustness to special reflective surfaces with a core of three designs.

Neural Video Depth Stabilizer(ICCV)

NVDS+: Towards Efficient and Versatile Neural Stabilizer for Video Depth Estimation(TPAMI)

Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian§, Guosheng Lin

Paper(ICCV) Paper(TPAMI) Github Project Page Dataset Page

• We propose a plug-and-play and bidirectional learning-based framework termed Neural Video Depth Stabilizer(NVDS), which can be directly adapted to different single-image depth predictors to remove flickers.
• We propose VDW dataset, which is currently the largest video depth dataset in the wild with the most diverse video scenes.

Diffusion-Augmented Depth Prediction with Sparse Annotations

Jiaqi Li, Yiran Wang§, Zihao Huang, Jinghong Zheng, Ke Xian, Zhiguo Cao, Jianming Zhang

The depth annotations collected by LiDAR in autonomous driving scenarios are highly sparse, and it is difficult for the model to reconstruct a dense and complete depth map from this learning. The previous methods have poor robustness in challenging scenarios such as night, rain, and dazzling light, and cannot meet the practical applications. We propose a plug-and-play framework based on diffusion modeling and object-guided integrality loss to enhance global and local structural integrity, respectively.