Spatio-Temporal Graph Scattering Transform

Authors: Chao Pan, Siheng Chen, Antonio Ortega

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now evaluate the performance of proposed ST-GST in skeleton-based action recognition task. ... We consider two datasets, MSR Action3D and NTU-RGB+D (cross-subject). ... Tables 1 and 2 compares the classification accuracies on MSR Action3D and NTU-RGB+D, respectively. We see that even without any training, the performance of ST-GST is better than other non-deep-learning and LSTM-based methods, and is comparable with state-of-the-art GCN-based methods in large-scale dataset.
Researcher Affiliation Collaboration Chao Pan University of Illinois at Urbana-Champaign Champaign, IL, USA chaopan2@illinois.edu; Siheng Chen Shanghai Jiao Tong University Shanghai, China sihengc@sjtu.edu.cn; Antonio Ortega University of Southern California Los Angeles, CA, USA antonio.ortega@ee.usc.edu; This work was mainly done while Chao Pan and Siheng Chen were working at Mitsubishi Electric Research Laboratories (MERL).
Pseudocode No The paper describes the proposed methods procedurally but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper does not include an unambiguous statement about releasing code for the described methodology or a direct link to a source-code repository.
Open Datasets Yes MSR Action3D dataset (Li et al., 2010) is a small dataset capturing indoor human actions. ... NTU-RGB+D (Liu et al., 2019) is currently the largest dataset with 3D joints annotations for human action recognition task.
Dataset Splits Yes Training and testing set is decided by cross-subject split for this dataset, with 288 samples for training and 269 for testing. ... The cross-subject benchmark of NTU-RGB+D includes 40,320 clips for training and 16,560 for testing.
Hardware Specification No The paper discusses computational efficiency but does not specify any particular hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies (e.g., library names with version numbers) needed to replicate the experiments.
Experiment Setup Yes The number of layers, L, the number of spatial wavelet scales, Js, and the number of temporal wavelet scales, Jt, are represented by (Js, Jt, L) for separable ST-GST, and (J, L) for joint ST-GST. ... Features output by ST-GST are then utilized by random forest classifier with 300 decision trees for classification. ... Geometric scattering wavelets are used in both domain, and the nonlinear activation σ( ) is absolute value function which has the property of energy-preserving.