ParkGaussian: Surround-view 3D Gaussian Splatting for Autonomous Parking

TL;DR

ParkGaussian addresses the challenge of 3D reconstruction for autonomous parking in GPS-denied underground garages by representing scenes with 3D Gaussian primitives and aligning reconstruction with downstream parking-slot perception. It introduces ParkRecon3D, a benchmark with over 40K synchronized surround-view frames and 60K parking-slot annotations, and couples 3D Gaussian Splatting with a differentiable surround-view IPM to render BEV while propagating perception-guided losses. A slot-aware reconstruction strategy leverages pretrained detectors (e.g., DMPR-PS and GCN-Parking) to concentrate geometric fidelity in slot regions through a mixed teacher–student weighting scheme and distribution alignment, improving both novel view synthesis and downstream slot detection. Experimental results show state-of-the-art reconstruction quality and stronger alignment with perception, highlighting the practical impact for autonomous parking systems and providing a valuable dataset and codebase for future research. The work thus offers a principled path toward perception-aware 3D reconstruction in challenging parking environments.

Abstract

Parking is a critical task for autonomous driving systems (ADS), with unique challenges in crowded parking slots and GPS-denied environments. However, existing works focus on 2D parking slot perception, mapping, and localization, 3D reconstruction remains underexplored, which is crucial for capturing complex spatial geometry in parking scenarios. Naively improving the visual quality of reconstructed parking scenes does not directly benefit autonomous parking, as the key entry point for parking is the slots perception module. To address these limitations, we curate the first benchmark named ParkRecon3D, specifically designed for parking scene reconstruction. It includes sensor data from four surround-view fisheye cameras with calibrated extrinsics and dense parking slot annotations. We then propose ParkGaussian, the first framework that integrates 3D Gaussian Splatting (3DGS) for parking scene reconstruction. To further improve the alignment between reconstruction and downstream parking slot detection, we introduce a slot-aware reconstruction strategy that leverages existing parking perception methods to enhance the synthesis quality of slot regions. Experiments on ParkRecon3D demonstrate that ParkGaussian achieves state-of-the-art reconstruction quality and better preserves perception consistency for downstream tasks. The code and dataset will be released at: https://github.com/wm-research/ParkGaussian

Figures (5)

• Figure 1: ParkRecon3D: a surround-view dataset for parking-aware reconstruction. It contains synchronized image streams from four fisheye cameras covering full 360° surroundings and four representative underground parking scenes. In total, the dataset includes 40K multi-camera sensor frames and 60K human-annotated parking-slot labels, supporting parking-slot perception and geometry-aware 3D reconstruction.
• Figure 2: Overview of the ParkGaussian pipeline. Four surround-view fisheye images are first rendered from 3D Gaussian primitives using UT-based projection for stable splatting under strong distortion. The rendered views are then passed through a differentiable IPM module to obtain a unified BEV map. A pretrained parking-slot detector (DMPR-PS or GCN-Parking) processes both rendered and ground-truth BEV maps to produce teacher–student structural guidance, from which slot-aware weights are constructed. These weights supervise reconstruction in both the IPM space and the camera-view space, forming the alignment and slot-aware objectives. ParkGaussian jointly optimizes Gaussian attributes toward photometric fidelity and perception-aligned slot geometry.
• Figure 3: Novel view synthesis visualization in 4 surround view fisheye images.
• Figure 4: Parking slot detection visualization in IPM images.
• Figure 5: Challenges in underground parking lots.