#3D perception

EmbodiedScan enhances embodied AI with a robust multi-modal 3D dataset, supporting effective visual grounding and scene interaction tasks in varied environments. The dataset offers over 5k scans, 1M ego-centric RGB-D views, and 160k categorized 3D boxes, bridging scene perception and language interaction. The Embodied Perceptron, a baseline framework, advances input processing for both structured tasks and real-world applications, with improvements such as dense semantic occupancy mapping and LVIS category compatibility.

PointTransformerV3 offers an efficient approach to 3D point cloud segmentation, providing improved speed and accuracy in semantic segmentation tasks on benchmarks like nuScenes and ScanNet. The project is continually updated in Pointcept v1.5, supplying valuable resources such as model weights and experiment records. Selected for oral presentation at CVPR'24, it utilizes Flash Attention to enhance computational efficiency and support scalable multi-dataset 3D representation learning.