ELMA: Energy-Based Learning for Multi-Agent Activity Forecasting

Authors: Yuke Li, Pin Wang, Lixiong Chen, Zheng Wang, Ching-Yao Chan1482-1490

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on two large-scale datasets prove that ELMA outperforms recent leading studies by an obvious margin.
Researcher Affiliation Academia Yu-Ke Li1, Pin Wang1, Li-Xiong Chen2, Zheng Wang3*, Ching-Yao Chan1* 1California PATH, UC Berkeley, 2Department of Engineering Science, University of Oxford, 3School of Computer Science, Wuhan University,
Pseudocode No The paper describes the algorithmic steps and processes like Langevin dynamics and MCMC but does not present them in a formalized pseudocode or algorithm block.
Open Source Code No The paper does not contain any statement about making the source code available or provide a link to a code repository.
Open Datasets Yes Two large-scale datasets, Activities in Extended Videos (Act EV/VIRAT) (Awad et al. 2018) benchmark and TITAN (Malla, Dariush, and Choi 2020), are used to assess the performance of ELMA.
Dataset Splits Yes TITAN contains 400 videos for training, 200 videos for validation, and 100 videos for test.
Hardware Specification Yes Our implementation uses Py Torch. The experiments are executed on four Nvidia Ge Force TITAN XPs, with 48 GB of memory in total.
Software Dependencies No The paper states 'Our implementation uses Py Torch' and mentions 'RMSProp optimizer' but does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes In our experiments, RMSProp optimizer (Goodfellow, Bengio, and Courville 2016) are employed with the learning rate initialized at 8 × 10−5. Our implementation uses Py Torch. The experiments are executed on four Nvidia Ge Force TITAN XPs, with 48 GB of memory in total. We observe his/her past 8 steps and forecast the activities of the subsequent 12 steps. Two stacked AG-CLSTM layers with 512 channels are leveraged to calculate Eq. 10. In practice, we consider building our graph with 50 nodes for the experiments.