Motion Graph Unleashed: A Novel Approach to Video Prediction

Authors: Yiqi Zhong, Luming Liang, Bohan Tang, Ilya Zharkov, Ulrich Neumann

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various datasets, including UCF Sports, KITTI and Cityscapes, highlight the strong representative ability of motion graph. Especially on UCF Sports, our method matches and outperforms the SOTA methods with a significant reduction in model size by 78% and a substantial decrease in GPU memory utilization by 47%.
Researcher Affiliation Collaboration Yiqi Zhong13 Luming Liang1 Bohan Tang2 Ilya Zharkov1 Ulrich Neumann3 1Microsoft 2University of Oxford 3University of Southern California 1{yiqizhong,lulian,zharkov}@microsoft.com 2bohan.tang@eng.ox.ac.uk 3{yiqizhon,uneumann}@usc.edu
Pseudocode No The paper contains figures depicting network architectures (e.g., Figure 7, 8, 9) but does not include any explicitly labeled “Pseudocode” or “Algorithm” blocks.
Open Source Code Yes Please refer to this link for the official code.
Open Datasets Yes For evaluation, we trained our video prediction pipeline in an end-to-end fashion on three public datasets: i) UCF Sports [36] (...); ii) KITTI [37] (...); and iii) Cityscapes [39] (i.e. 3,475 driving videos with 2,945 in the training set and 500 in the validation set).
Dataset Splits Yes On the UCF Sports MMVP split, the validation dataset has been divided into three categories: the easy (SSIM 0.9), intermediate (0.6 SSIM < 0.9), and hard subsets (SSIM < 0.6), which take up 66%, 26%, and 8% of the full set respectively. (...) iii) Cityscapes [39] (i.e. 3,475 driving videos with 2,945 in the training set and 500 in the validation set).
Hardware Specification Yes We implement the video prediction system using Py Torch [54] and conduct end-to-end training on a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions software like “PyTorch [54]”, “AdamW optimizer [55]”, and “cosine decay scheduler [56]” but does not provide specific version numbers for these software components.
Experiment Setup Yes In Table 9, we demonstrate the hyper-parameter setting for each dataset. Table 9 includes specific values for Image feat., Tendency feat., Location feat., Number of Graph views, k, Epoch, and Loss for UCF Sports, Cityscapes, and KITTI.