Conceptual
Browse Instruct Curate

PanopticRecon++: Cross-Attention Open-Vocabulary Panoptic 3D Reconstruction

An end-to-end open-vocabulary panoptic 3D reconstruction method that reformulates the task as cross-attention: learnable 3D Gaussians serve as instance queries attending over the scene's 3D embedding field, injecting spatial priors while keeping queries and field jointly optimizable. Aligns 2D instance IDs across frames by optimal linear assignment on rendered masks and fuses instance/semantic probabilities in a panoptic head under a panoptic loss, with query-token count adapting to object count.