Recently, the acceptance results of The 40th Annual AAAI Conference on Artificial Intelligence (AAAI 2026) were announced. The paper titled "The Structure-Equivalent Prior: Unifying Temporal Dynamics and 3D Evolution in 4D Latent Space" by the research team led by Professor Wang Kunfeng from our school was successfully accepted. The first author of this paper is Gao Jingyuan, a 2023-level Master’s candidate of the School of Information Science and Technology. The co-corresponding authors are Associate Professor Shen Tianyu (collaborative supervisor) and Professor Wang Kunfeng (graduate supervisor). Beijing University of Chemical Technology is the sole completing institution.
Addressing the limitations of dynamic 3D scene reconstruction methods in modeling spatiotemporal consistency, this study proposes a unified 4D latent space representation method (SEP-4D) based on the Structure-Equivalent Prior. Existing methods usually simply combine temporal and spatial dimensions when processing dynamic 3D scenes, ignoring the inherent correlation between spatiotemporal evolution, which leads to reconstruction results lacking physical rationality and structural coherence. In addition, encoding methods based on discretized Tokenizers struggle to model fine-grained structural time-varying details, thus resulting in suboptimal reconstruction performance. By introducing the Structure-Equivalent Prior, SEP-4D models temporal evolution as continuous geometric deformation of spatial structures, thereby achieving high-quality spatiotemporally consistent reconstruction in a unified 4D latent space. This method employs a plane decomposition encoder to decompose dynamic scenes into multiple learnable 2D feature planes, and explicitly constrains temporal changes to originate from smooth deformation of spatial structures through an inter-plane spatiotemporal fusion mechanism based on probabilistic fusion. Experimental results demonstrate that SEP-4D significantly outperforms state-of-the-art methods in 4D occupancy grid reconstruction tasks, and maintains excellent accuracy and stability especially in long-sequence scenarios.
The AAAI Conference on Artificial Intelligence (AAAI) is one of the top-tier international academic conferences in the field of artificial intelligence. Founded in 1979, it is hosted by the Association for the Advancement of Artificial Intelligence. According to the latest global academic journal and conference influence rankings released by Google Scholar, AAAI ranks 4th in the "Artificial Intelligence" category with an h5-index of 232, and is recognized as a Class A top international academic conference in the field of artificial intelligence by the China Computer Federation (CCF). AAAI-26 received a total of 23,680 valid submissions, with 4,167 papers finally accepted, resulting in an acceptance rate of 17.6%. The acceptance of this paper marks another breakthrough made by our school in top-tier international conferences in the field of artificial intelligence.
This work was supported by grants from the National Natural Science Foundation of China and the Beijing Natural Science Foundation.