A Coarse-to-Fine Approach to Multi-Modality 3D Occupancy Grounding
This work addresses the need for more precise object perception in spatially aware systems like autonomous driving, though it is incremental as it builds on existing datasets and methods.
The paper tackles the problem of inaccurate object representation in visual grounding for autonomous driving by introducing a benchmark for 3D occupancy grounding that integrates natural language with voxel-level annotations, and proposes GroundingOcc, an end-to-end model that outperforms existing baselines on this task.
Visual grounding aims to identify objects or regions in a scene based on natural language descriptions, essential for spatially aware perception in autonomous driving. However, existing visual grounding tasks typically depend on bounding boxes that often fail to capture fine-grained details. Not all voxels within a bounding box are occupied, resulting in inaccurate object representations. To address this, we introduce a benchmark for 3D occupancy grounding in challenging outdoor scenes. Built on the nuScenes dataset, it integrates natural language with voxel-level occupancy annotations, offering more precise object perception compared to the traditional grounding task. Moreover, we propose GroundingOcc, an end-to-end model designed for 3D occupancy grounding through multi-modal learning. It combines visual, textual, and point cloud features to predict object location and occupancy information from coarse to fine. Specifically, GroundingOcc comprises a multimodal encoder for feature extraction, an occupancy head for voxel-wise predictions, and a grounding head to refine localization. Additionally, a 2D grounding module and a depth estimation module enhance geometric understanding, thereby boosting model performance. Extensive experiments on the benchmark demonstrate that our method outperforms existing baselines on 3D occupancy grounding. The dataset is available at https://github.com/RONINGOD/GroundingOcc.