The field of 3D vision and scene understanding is rapidly advancing, with a focus on developing more efficient and accurate methods for 3D object detection, segmentation, and steering estimation. A key trend in this area is the integration of 2D and 3D data, leveraging the strengths of both to improve performance. Another area of innovation is the development of semi-supervised and open-vocabulary detection methods, which aim to reduce the need for large amounts of labeled data. These advancements have the potential to improve a range of applications, from autonomous driving to orthodontic analysis.

Noteworthy papers in this area include:

• Seg2Box, which proposes a novel method for 3D object detection using semantic labels, achieving significant improvements in performance.
• GeoT, which introduces a geometry-guided instance-dependent transition matrix for semi-supervised tooth point cloud segmentation, demonstrating comparable performance to fully supervised methods with limited labeled data.
• DINO in the Room, which leverages 2D foundation models for 3D segmentation, achieving state-of-the-art results on indoor and outdoor benchmarks.
• GLRD, which proposes a global-local collaborative reason and debate framework for 3D open-vocabulary detection, showing superiority over existing methods in partial and full open-vocabulary settings.