PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders

Authors: Xiangdong Zhang, Shaofeng Zhang, Junchi Yan

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments including object classification, few-shot learning and segmentation to demonstrate the superior performance of our method over Point-MAE.
Researcher Affiliation Academia Xiangdong Zhang , Shaofeng Zhang , Junchi Yan Dept. of CSE & School of AI & Mo E Key Lab of Al, Shanghai Jiao Tong University {zhangxiangdong, sherrylone, yanjunchi}@sjtu.edu.cn
Pseudocode No The paper describes algorithms and modules in prose and through diagrams (Figure 2) but does not include a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes The code is available at https://github.com/a Hap Bean/PCP-MAE .
Open Datasets Yes PCP-MAE is pre-trained on Shape Net [3] which consists of about 51,300 clean 3-D models, covering 55 common object categories.
Dataset Splits No The paper describes various datasets used for pre-training and fine-tuning (ShapeNet, Scan Object NN, ModelNet40, ShapeNet Part, S3DIS) and mentions augmentation and patching strategies, but it does not provide explicit numerical training, validation, or test splits (e.g., '80/10/10 split') in its text. It relies on 'established protocols' for these datasets.
Hardware Specification Yes We run all experiments with single GPU either using RTX 3090 (24GB) or V100 (32GB).
Software Dependencies No The paper mentions 'Adam W optimizer [22]' and 'A cosine learning rate decay scheduler [21]' but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes The backbone of our pre-trained PCP-MAE consists of standard Transformer blocks where the encoder has 12 Transformer blocks and the decoder has 4, aligned with Point MAE [24]. The hidden dimension of Transformer blocks is 384 and the number of heads is 6. ... The PCP-MAE is pre-trained for 300 epochs using an Adam W optimizer [22] with a batch size of 128. The initial learning rate is set at 0.0005, with a weight decay of 0.05. A cosine learning rate decay scheduler [21] is utilized.