Multi-View Intelligent Perception多视角智能感知研究组

Research Interests

This overarching theme is driven by two main pillars of research:

Pixel-level Vision Tasks

This track addresses the challenge of restoring high-fidelity visual information when data is corrupted by environmental degradation (e.g., low light, rain, reflections, and sensor noise). Low-Light Enhancement & Denoising: Exploring view-consistency priors and multi-stream progressive networks to achieve scene-adaptive illumination and noise reduction. Interference Removal (Reflection & Raindrops): Innovating hierarchical multi-plane image (MPI) construction and multi-layer interaction mechanisms to achieve pixel-level decoupling of background and interference layers. Light Field Super-Resolution: Combining flexible hybrid lenses and structure-aware neural rendering to reconstruct high-resolution, high-quality light field images.

Image Super-Resolution

Generating high-resolution images from multiple low-resolution views using deep learning algorithms to surpass hardware limitations.

Occlusion Removal

Analyzing the spatial relationship between occluders and background using multi-view information to intelligently remove foreground occlusions.

Reflection Separation

Separating reflection and transmission layers by leveraging multi-view consistency constraints for images affected by glass reflections.

Image Deblurring

Restoring image clarity from motion blur or defocus blur through the design of specialized deep neural networks.

Denoising / Deraining

Enhancing visual quality in noisy or adverse weather conditions using multi-view information fusion and noise modeling.

Low-light Enhancement

Improving brightness and visual quality in low-light scenes while maintaining view consistency by utilizing complementary scene details.

3D Reconstruction Tasks

This track tackles the challenges of estimating geometry and understanding semantics in the presence of severe occlusions, weak textures, and complex materials. Unsupervised & Occlusion-Robust Depth Estimation: Proposing epipolar consistent attention aggregation networks to overcome physical occlusion bottlenecks and achieve highly accurate, unsupervised light field disparity estimation. Geometry Perception for Complex Materials: Solving depth ambiguity issues for transparent and reflective surfaces through dual-layer depth estimation and decoupling strategies. Semantic Understanding & Occlusion Removal: Integrating structural priors into LF semantic segmentation and advancing flexible 3D occlusion mask learning within Neural Radiance Fields (NeRF).

Depth Estimation

Accurately estimating scene depth distribution using epipolar plane image analysis and deep learning algorithms from multi-view inputs.

View Synthesis / Novel View Synthesis

Reconstructing 3D scene models from sparse view inputs and synthesizing realistic novel views using techniques like NeRF.

Salient Object Detection & Semantic Segmentation

Detecting salient objects and performing pixel-level semantic segmentation by combining depth information with appearance features.

Salient Object Detection Semantic Segmentation Scene Understanding

Publications

Representative papers published in top-tier international journals and conferences

Pixel-level Vision Tasks

3D Reconstruction Tasks

Collaborative Research