Wavelet-Driven Spatiotemporal Predictive Learning: Bridging Frequency and Time Variations

Authors: Xuesong Nie, Yunfeng Yan, Siyuan Li, Cheng Tan, Xi Chen, Haoyuan Jin, Zhihang Zhu, Stan Z. Li, Donglian Qi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across various real-world scenarios, such as driving scene prediction, traffic flow prediction, human motion capture, and weather forecasting, demonstrate that our proposed Wa ST achieves state-of-the-art performance over various spatiotemporal prediction methods.
Researcher Affiliation Academia Xuesong Nie1, Yunfeng Yan1*, Siyuan Li1, Cheng Tan1, Xi Chen3, Haoyuan Jin1, Zhihang Zhu1, Stan Z. Li2, Donglian Qi1 1Zhejiang University, Zhejiang, China 2School of Engineering, Westlake University, Zhejiang, China 3Department of Computer Science, The University of Hong Kong, Hong Kong, China
Pseudocode No The paper provides architectural diagrams and mathematical formulations but no pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/xuesongnie/Wa ST.
Open Datasets Yes Dataset statistics are summarized in Table 1. Table 1: The detailed summary of the dataset statistic. The number of samples, input frames T, and predicted frames T are shown for the training and testing sets. (Lists Kitti&Caltech, Human3.6M, Taxi BJ, Weather Bench, which are standard public datasets.)
Dataset Splits No Table 1 provides train and test set sizes, but no explicit validation set split or size is mentioned.
Hardware Specification Yes Our proposed method is implemented in Pytorch and conducts experiments on a single NVIDIA-V100 GPU.
Software Dependencies No Our proposed method is implemented in Pytorch and conducts experiments on a single NVIDIA-V100 GPU. (Only mentions "Pytorch" without a specific version number, which is insufficient for reproducibility.)
Experiment Setup Yes The model trained with a minibatch of 16 video sequences, employs the Adam W optimizer, One Cycle learning rate scheduler, and weight decay of 5e 2. Optimal learning rate is chosen from 1e 2, 5e 3, 1e 3 for stable training. We utilize stochastic depth for regularization to avoid overfitting.