Omni-Query Active Learning for Source-Free Domain Adaptive Cross-Modality 3D Semantic Segmentation

Authors: Jianxiang Xie, Yao Wu, Yachao Zhang, Zhongchao Shi, Jianping Fan, Yuan Xie, Yanyun Qu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the effectiveness of our proposed method, achieving up to 99.64% of fully supervised performance with only 3% of labels, and consistently outperforming comparison methods across various scenarios. We utilize four public datasets: nu Scenes-Lidarseg(Caesar et al. 2020), Virtual KITTI (Gaidon et al. 2016), Semantic KITTI(Behley et al. 2019), and A2D2(Geyer et al. 2020), to construct four domain adaptation experimental scenarios to validate our method.
Researcher Affiliation Collaboration 1School of Informatics, Xiamen University 2Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University 3School of Computer Science and Technology, East China Normal University 4Lenovo Research
Pseudocode Yes A detailed algorithm can be found in supplementary material.
Open Source Code Yes Code https://github.com/Kylin-XJX/Active SFDA
Open Datasets Yes We utilize four public datasets: nu Scenes-Lidarseg(Caesar et al. 2020), Virtual KITTI (Gaidon et al. 2016), Semantic KITTI(Behley et al. 2019), and A2D2(Geyer et al. 2020)
Dataset Splits No Detailed information is provided in the supplementary material.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are mentioned in the paper.
Software Dependencies No The paper mentions architectural components like U-Net, Res Net-34, and Sparse Conv Net, but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We voxelize the point cloud to a size of 5cm... We initialize the target model using the source model trained only on the source domain... except for the Virtual KITTI Semantic KITTI scenario, which runs 30k iterations to prevent overfitting, all other scenarios are trained for 100k iterations... We experiment under various label budgets p and set the initial percent b of the budget to 50... The overall loss consists of three parts: active label loss, pseudo-label loss, and cross-modal loss.