STDiff: Spatio-Temporal Diffusion for Continuous Stochastic Video Prediction

Authors: Xi Ye, Guillaume-Alexandre Bilodeau

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluated the performance of the proposed STDiff model on KITTI (Geiger et al. 2013), Cityscapes (Cordts et al. 2016), KTH (Schuldt, Laptev, and Caputo 2004), BAIR (Ebert et al. 2017) and stochastic moving MNIST (SMMNIST) (Denton and Fergus 2018) datasets.
Researcher Affiliation Academia LITIV, Polytechnique Montr eal xi.ye@polymtl.ca, gabilodeau@polymtl.ca
Pseudocode Yes Algorithm 1: Training
Open Source Code Yes Our code is available at https: //github.com/Xi Ye20/STDiff Project.
Open Datasets Yes We evaluated the performance of the proposed STDiff model on KITTI (Geiger et al. 2013), Cityscapes (Cordts et al. 2016), KTH (Schuldt, Laptev, and Caputo 2004), BAIR (Ebert et al. 2017) and stochastic moving MNIST (SMMNIST) (Denton and Fergus 2018) datasets.
Dataset Splits No The number of past frames and future frames to predict is determined according to the experimental protocols of previous works (see appendix in the ar Xiv version of this paper for more details (Ye and Bilodeau 2023a)).
Hardware Specification Yes All models are trained with 4 NVIDIA V100 Volta GPU (32G memory).
Software Dependencies No No specific software dependencies with version numbers were mentioned.
Experiment Setup No The paper describes some general training procedures like random sampling of future time steps and the number of predictions/sampling steps, but it does not provide specific hyperparameters such as learning rate, batch size, or optimizer settings in the main text.