Enhancing Generalizability via Utilization of Unlabeled Data for Occupancy Perception

Authors: Ruihang Li, Tao Li, Shanding Ye, Kaikai Xiao, Zheng Huangnan, Zhe Yin, Zhijie Pan

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiment results on two challenging datasets nu Scenes and Waymo, demonstrate that our method not only achieves state-of-the-art generalizability but also enhances the model s perceptual capabilities within the source domain by utilizing unlabeled data. Quantitative Experiments and Ablation Studies
Researcher Affiliation Academia Ruihang Li, Tao Li, Shanding Ye, Kaikai Xiao, Zheng Huangnan, Zhe Yin, Zhijie Pan College of Computer Science and Technology, Zhejiang University Hangzhou, 310027 China EMAIL
Pseudocode No The paper describes methods and provides equations, but it does not contain any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Additionally, we have open-sourced our code, contributing to the community.
Open Datasets Yes Our experiments use the nu Scenes and Waymo datasets (Caesar et al. 2020; Sun et al. 2020) with OCC labels from OCC-nu Scenes and OCC-Waymo.
Dataset Splits No The paper defines 'nu Scenes as the source domain and Waymo as the target domain' for benchmarks, and vice-versa, but does not provide specific training/validation/test splits (e.g., percentages or sample counts) for these datasets.
Hardware Specification Yes Experiments run on six 4090 devices
Software Dependencies No The paper mentions using Res Net and Adam W optimizer, but does not provide specific version numbers for any software libraries or dependencies (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes Methods are trained for 24 epochs in the source domain using the Adam W optimizer with a learning rate of 7 x 10^-5. For cross-domain learning, we pretrain self-supervised depth estimation and Semantic Query Adversarial Fusion, setting their backpropagation ratios to 0.1. Confidence and smoothness thresholds are 0.85 and 0.8, respectively. Experiments run on six 4090 devices, with a learning rate of 3.5 x 10^-5, continuing until the pseudo-label selection ratio exceeds 35%.