Chaolei Wang

We propose VoxelFlow, a novel framework that effectively combines 2D mask semantics with 3D geometric features. Our core insight is to leverage the implicit semantic and spatial information embedded in 2D masks, using them to replace rigid 3D priors.

We propose splitting and growing reliable semantic mask for high-fidelity 3D instance segmentation (SGS-3D), a novel "split-then-grow" framework that first purifies and splits ambiguous lifted masks using geometric primitives, and then grows them into complete instances within the scene.

We introduce Vireo, a novel single stage framework for OV-DGSS that unifies the strengths of OVSS and DGSS for the first time. Vireo builds upon frozen Visual Foundation Models (VFMs) and incorporates scene geometry via Depth VFMs to extract domain-invariant structural features.

This work develops CLIP4Geo, a unified GeoFM that integrates satellite imagery, LiDAR point clouds, geo-tagged points of interest (POIs), and textual descriptions into a shared representation space.

We propose a framework named CityInsight for analyzing urban building morphology from remote sensing imagery. First, we establish a semantic segmentation network, dual-condition diffusion network (DC-Net), based on a diffusion model to accurately identify building footprints from remote sensing images. Second, we use uncertainty attention and condition attention to generate spatial and semantic priors. Finally, we design acondition injection module to incorporate spatial and semantic information into the diffusion learning.

We propose an effective individual tree segmentation method capable of accurately extracting single trees in urban and forest scenes. The proposed approach consists of two primary steps: (1) We design the Semantic-Driven Instance Clustering to combine the semantic prior with the instance embeddings. (2) We introduce the Online Semantic Clustering for intra-class potential semantic discriminability, improving the instance representation within the same semantic class.

3D scene flow aims to recover the dense geometry and 3D motion of dynamic scenes. This paper explores the transformation and adaptation of the 2D-3D feature space in the joint estimation of optical flow and scene flow. Our key insight is to fully leverage the unique characteristics of each modality and maximize their inter-modality complementarity. To achieve this, we propose a novel architecture, named PAFlow, which consists of Camera-LiDAR Adaptation and Spatial Characteristics Adaptation.

To address these issues, we aim to propose an individual tree segmentation algorithm applicable for complex mangrove scenes.“Method” This study innovatively combines deep learning and traditional algorithms to propose a high-precision individual tree segmentation framework for TLS point clouds in complex mangrove scenes. The framework initially employs the deep learning network RandLA-Net for ground filtering and wood-leaf separation. Subsequently, mangrove main stems are segmented using a connected component segmentation method. Finally, individual tree segmentation is achieved through the multiple tree tops constraint module.“Results” To assess the accuracy of the algorithm, we use three measures: completeness, correctness, and accuracy. We also conduct a comparative analysis with two classical algorithms.