XMask3D++: Cross-Modal Mask Reasoning for Open Vocabulary 3D Segmentation

Existing methodologies in open vocabulary 3D semantic segmentation primarily concentrate on establishing a unified feature space encompassing 3D, 2D, and textual modalities. Nevertheless, traditional techniques such as global feature alignment or vision-language model distillation tend to impose only approximate correspondence, struggling notably with delineating fine-grained segmentation boundaries. To address this gap, we propose a more meticulous mask-level alignment between 3D features and the 2D-text embedding space through a cross-modal mask reasoning framework, XMask3D. In our approach, we developed a mask generator based on the denoising UNet from a pre-trained diffusion model, leveraging its capability for precise textual control over dense pixel representations and enhancing the open-world adaptability of the generated masks. We integrate 3D global features as implicit conditions into the pre-trained 2D denoising UNet, enabling the generation of segmentation masks with additional 3D geometry awareness. Subsequently, the generated 2D masks are employed to align mask-level 3D representations with the vision-language feature space, thereby augmenting the open vocabulary capability of 3D geometry embeddings. Finally, we fuse complementary 2D and 3D mask features, resulting in competitive performance across multiple benchmarks for 3D open vocabulary semantic segmentation. To improve the quality of back-projected 3D masks, we further introduce a 2D-to-3D mask adaptation module that iteratively refines both 3D mask contours and condenses mask prototype features in the advanced XMask3D++ framework. The enhanced mask prototypes, which are also explicitly regularized by the mask loss, are then integrated into the fused features to boost open-vocabulary capacity. We demonstrate the effectiveness of XMask3D++ through consistent improvements across various benchmarks and datasets, covering both indoor and outdoor scenarios. We further conduct zero-shot and domain transfer experiments to validate the generalization capability of the proposed method. In addition, we evaluate the framework's flexibility by replacing the 3D geometric feature extractor and the 2D open-mask generator with more advanced models, demonstrating the scalability and adaptability of our approach.