SE(3)-bi-equivariant Transformers for Point Cloud Assembly

Authors: Ziming Wang, Rebecka Jörnsten

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally show the effectiveness of BITR in practical tasks.
Researcher Affiliation Academia Ziming Wang, Rebecka Jörnsten Department of Mathematical Sciences University of Gothenburg and Chalmers University of Technology zimingwa@chalmers.se, jornsten@chalmers.se
Pseudocode No The paper describes algorithms and formulations but does not include distinct pseudocode blocks or figures labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes Code is available at: https://github.com/wzm2256/Bi Tr
Open Datasets Yes We train BITR on the bunny shape [33]. ... We evaluate BITR on assembling the shapes in Shape Net [6], BB dataset [31], 7Scenes [32] and ASL [22].
Dataset Splits Yes The airplane dataset used in Sec. 6.3.1 contains 715 random training samples and 103 random test samples. The wine bottle dataset used in Sec. 6.4 contains 331 training and 41 test samples. We adopt the few-shot learning setting in the manipulation tasks in Sec. 6.6: we use 30 training and 5 test samples for mug-hanging; we use 40 training samples and 10 test samples for bowl-placing.
Hardware Specification Yes We run all experiments using a Nvidia T4 GPU card with 16G memory.
Software Dependencies No The paper mentions 'Adam optimizer [15]', 'Open3D [46]', and 'Py Bullet [8]' but does not provide specific version numbers for these or other key software libraries like PyTorch or CUDA, which are necessary for reproducible software dependencies.
Experiment Setup Yes We extract L = 32 key points for each PC. The SE(3)-transformer and the SE(3) SE(3)-transformer both contain 2 layers with c = 4 channels. We consider k = 24 nearest neighborhoods for message passing. We train BITR using Adam optimizer [15] with learning rate 1e 4. We use the loss function L = r T rgt I 2 2 + tgt t 2 2