Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding
Authors: LinshuangDiao, Sensen Song, Yurong Qian, Dayong Ren
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our pre-training Zigzag Point Mamba weights significantly boost downstream tasks, achieving a 1.59% m Io U gain on Shape Net Part for part segmentation, a 0.4% higher accuracy on Model Net40 for classification, and 0.19%, 1.22%, and 0.72% higher accuracies respectively for the classification tasks on the OBJ-BG, OBJ-ONLY, and PB-T50-RS subsets of Scan Object NN. |
| Researcher Affiliation | Academia | 1Key Laboratory of Signal Detection and Processing, Xinjiang University 2Joint International Research Laboratory of Silk Road Multilingual Cognitive Computing, Xinjiang University 3National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China |
| Pseudocode | Yes | Algorithm 1 Semantic-Siames Masking Strategy |
| Open Source Code | Yes | Code is available at https://github.com/Rabbitttttt218/Zigzag Point Mamba. |
| Open Datasets | Yes | Our pre-training Zigzag Point Mamba weights significantly boost downstream tasks, achieving a 1.59% m Io U gain on Shape Net Part for part segmentation, a 0.4% higher accuracy on Model Net40 for classification, and 0.19%, 1.22%, and 0.72% higher accuracies respectively for the classification tasks on the OBJ-BG, OBJ-ONLY, and PB-T50-RS subsets of Scan Object NN. |
| Dataset Splits | No | The paper mentions using well-known datasets like ModelNet40, ShapeNet Part, and Scan Object NN, which typically have standard splits. However, it does not explicitly state the specific training/test/validation splits used for these datasets or provide percentages/sample counts for general experimentation. It only describes sampling for few-shot learning. For example: "Model Net40: In the classification experiment on Model Net40 [31]. As shown in Table 2, we conducted experiments without using the voting strategy." and "The Model Net Few Shot dataset, constructed based on Model Net40, is specifically designed for few-shot learning research. The selection of few-shot data in it strictly follows the following rules: randomly select 5 or 10 categories from the 40 categories of Model Net40 as the task categories, and then randomly select 10 or 20 samples from each selected category as the support set." |
| Hardware Specification | Yes | Experiments are run on a NVIDIA A40 GPU. |
| Software Dependencies | No | The paper mentions the use of 'Adam W optimizer' and 'Cosine annealing scheduler' which are algorithms, but does not provide specific version numbers for any software libraries (e.g., PyTorch, TensorFlow) or CUDA. |
| Experiment Setup | Yes | Training uses Adam W optimizer (lr = 0.001), Cosine annealing scheduler, 300 epochs, batch size 128, and loss of Chamfer distance L2 [4]. Experiments are run on a NVIDIA A40 GPU. We randomly select one path, setting the random masking ratio at 0.6 and SMS ratio at 0.8. |