Explore In-Context Learning for 3D Point Cloud Understanding

Authors: Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, Mengyuan Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to validate the versatility and adaptability of our proposed methods in handling a wide range of tasks.
Researcher Affiliation Academia 1Sun Yat-sen University 2S-Lab, Nanyang Technological University 3Department of Computer Science, ETH Zurich 4Key Laboratory of Machine Perception, Shenzhen Graduate School, Peking University
Pseudocode No The paper describes its methods through text and diagrams, but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes https://github.com/fanglaosi/Point-In-Context
Open Datasets Yes Firstly, we obtain samples from publicly available datasets, such as Shape Net [7], Shape Net Part [42]
Dataset Splits No The paper mentions training and testing but does not explicitly provide details about a distinct validation set split or its size/percentage for reproduction.
Hardware Specification Yes Test speed ... tested on one NVIDIA RTX 3080 Ti GPU.
Software Dependencies No The paper mentions the use of 'Adam W optimizer' and 'standard transformer', but does not specify versions of programming languages, libraries, or frameworks used for implementation, such as PyTorch or TensorFlow versions.
Experiment Setup Yes We sample 1024 points of each point cloud and divide it into N = 64 point patches, each with M = 32 neighborhood points. We set the mask ratio as 0.7. For PIC-Sep, we merge the feature of input and target at the third block. We randomly select a prompt pair that performs the same task with the query point cloud from the training set. We use an Adam W optimizer [23] and cosine learning rate decay, with the initial learning rate as 0.001 and a weight decay of 0.05. All models are trained for 300 epochs.