Adaptive Wavelet Transformer Network for 3D Shape Representation Learning

Authors: Hao Huang, Yi Fang

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that AWT-Net achieves competitive performance on 3D shape classification and segmentation benchmarks.
Researcher Affiliation Academia Hao Huang, Yi Fang NYU Multimedia and Visual Computing Lab, USA NYUAD Center for Artificial Intelligence and Robotics (CAIR), Abu Dhabi, UAE NYU Tandon School of Engineering, New York University, USA New York University Abu Dhabi, UAE {hh1811,yfang}@nyu.edu
Pseudocode Yes Algorithm 1: Even-Odd Node Split
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes We evaluate our model on 3D shape classification and part segmentation tasks on three datasets following the settings (Qi et al., 2017a;b; Wang et al., 2019; Xu et al., 2021b; Zhao et al., 2021): Model Net40 (Wu et al., 2015), Scan Object NN (Uy et al., 2019) and Shape Net Part (Yi et al., 2016).
Dataset Splits Yes We evaluate our model on 3D shape classification and part segmentation tasks on three datasets following the settings (Qi et al., 2017a;b; Wang et al., 2019; Xu et al., 2021b; Zhao et al., 2021): Model Net40 (Wu et al., 2015) which contains 9,843 training shapes and 2,468 test shapes across 40 categories.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU model, CPU type, memory) used to run its experiments. It only mentions 'The computational overhead is measured in second for a single forward and backward pass.'
Software Dependencies No We implement our model in Py Torch (Paszke et al., 2019). (No version number for PyTorch or any other software is provided.)
Experiment Setup Yes We use SGD optimizer with momentum and weight decay set to 0.9 and 0.0001, respectively. For both 3D shape classification and 3D object part segmentation tasks, we train for 350 epochs. The initial learning rate is set to 0.2 for shape classification and 0.005 for object part segmentation, and clipped at 0.9e 5. The values of λ1 and λ2 are set to 0.1 for all tasks.