Memory-Augmented Re-Completion for 3D Semantic Scene Completion

Authors: Yu-Wen Tseng, Sheng-Ping Yang, Jhih-Ciang Wu, I-Bin Liao, Yung-Hui Li, Hong-Han Shuai, Wen-Huang Cheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on the SSCBench-KITTI-360 and Semantic KITTI datasets validate the effectiveness of our approach.
Researcher Affiliation Collaboration 1 National Taiwan University, Taiwan 2 National Yang Ming Chiao Tung University, Taiwan 3 National Taiwan Normal University, Taiwan 4 Hon Hai Research Institute, Taiwan
Pseudocode Yes Algorithm 1: Memory Updating
Open Source Code Yes Code https://github.com/ywtseng0226/MARE
Open Datasets Yes The evaluation is performed on SSCBench-KITTI-360 (Li et al. 2024) and Semantic KITTI (Behley et al. 2019) datasets
Dataset Splits Yes SSCBench-KITTI-360 provides 9 video sequences, with 7 for training, 1 for validation, and 1 for testing. Semantic KITTI provides 20 video sequences, with 9 for training, 1 for validation, and 11 for testing.
Hardware Specification Yes trained in an end-to-end manner with two NVIDIA V100 GPUs for 30 epochs
Software Dependencies No The paper mentions optimizers (Adam W), backbones (Res Net-50), and encoders (Mask DINO) but does not provide specific version numbers for software libraries or programming languages like Python or PyTorch.
Experiment Setup Yes The model is trained in an end-to-end manner with two NVIDIA V100 GPUs for 30 epochs, with a batch size of two images. We employ the Adam W as the optimizer, Res Net-50 as the backbone, and the pretrained weights of Mask DINO as the tokenbased Encoder. In the Regional Memory Bank, we set the size |B| as 1024 and the number of neighbor tokens k as 3. In the Re-completion pipeline, we re-complete the scene for two iterations.