SuperOcc: Toward Cohesive Temporal Modeling for Superquadric-based Occupancy Prediction

TL;DR

SuperOcc introduces a cohesive temporal modeling framework for superquadric-based 3D occupancy prediction in autonomous driving. By unifying view-centric and object-centric temporal cues, it leverages historical image features and query propagation to improve temporal consistency. A multi-superquadric decoding strategy enhances geometric expressiveness without sacrificing query sparsity, and a tile-based splatting operator significantly boosts computational efficiency. Extensive experiments on Occ3D and SurroundOcc demonstrate state-of-the-art performance and favorable efficiency, validating the method's practicality for real-time perception. The work advances sparse, geometry-rich scene representations for robust occupancy prediction in dynamic driving environments.

Abstract

3D occupancy prediction plays a pivotal role in the realm of autonomous driving, as it provides a comprehensive understanding of the driving environment. Most existing methods construct dense scene representations for occupancy prediction, overlooking the inherent sparsity of real-world driving scenes. Recently, 3D superquadric representation has emerged as a promising sparse alternative to dense scene representations due to the strong geometric expressiveness of superquadrics. However, existing superquadric frameworks still suffer from insufficient temporal modeling, a challenging trade-off between query sparsity and geometric expressiveness, and inefficient superquadric-to-voxel splatting. To address these issues, we propose SuperOcc, a novel framework for superquadric-based 3D occupancy prediction. SuperOcc incorporates three key designs: (1) a cohesive temporal modeling mechanism to simultaneously exploit view-centric and object-centric temporal cues; (2) a multi-superquadric decoding strategy to enhance geometric expressiveness without sacrificing query sparsity; and (3) an efficient superquadric-to-voxel splatting scheme to improve computational efficiency. Extensive experiments on the SurroundOcc and Occ3D benchmarks demonstrate that SuperOcc achieves state-of-the-art performance while maintaining superior efficiency. The code is available at https://github.com/Yzichen/SuperOcc.

Figures (3)

• Figure 1: Comparison of speed-accuracy trade-offs of different methods on (a) Occ3D and (b) SurroundOcc benchmarks. SuperOcc achieves the optimal trade-off, delivering high accuracy while maintaining efficient inference.
• Figure 2: Framework of SuperOcc. Given multi-view image sequences, SuperOcc constructs a superquadric-based sparse scene representation for 3D occupancy prediction. To achieve comprehensive temporal modeling, the framework extracts fine-grained spatio-temporal context through the interaction of sparse queries with multi-frame image features, while efficiently exploiting informative historical priors via query propagation. Furthermore, each updated query is decoded into a set of semantic superquadrics through a multi-superquadric decoding strategy. Finally, voxel-level occupancy prediction is generated through an efficient superquadric-to-voxel splatting process.
• Figure 3: Qualitative results of SuperOcc on the SurroundOcc wei2023surroundocc benchmark. For each scene, the predicted superquadrics, resulting occupancy prediction, and ground-truth occupancy are shown. Colors indicate semantic categories for both superquadrics and voxels.