SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents

Authors: Wei Xiang, Haoteng YIN, He Wang, Xiaogang Jin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on two public benchmarks including 25 scenes demonstrate that Social CVAE significantly improves prediction accuracy compared with the state-of-the-art methods, with up to 16.85% improvement in Average Displacement Error (ADE) and 69.18% improvement in Final Displacement Error (FDE). We conduct extensive experiments on two popular benchmark datasets (ETH-UCY (Pellegrini et al. 2009; Lerner, Chrysanthou, and Lischinski 2007) and SDD (Robicquet et al. 2016)), and demonstrate Social CVAE s superiority over existing state-of-the-art methods in terms of prediction accuracy.
Researcher Affiliation Academia Wei Xiang1*, Haoteng Yin2*, He Wang3, Xiaogang Jin1 1State Key Lab of CAD&CG, Zhejiang University 2Department of Computer Science, Purdue University 3Department of Computer Science, University College London xiangwvivi@gmail.com, yinht@purdue.edu, he wang@ucl.ac.uk, jin@cad.zju.edu.cn
Pseudocode No The paper describes the methodology in text and uses diagrams, but it does not contain a structured pseudocode or algorithm block explicitly labeled as such.
Open Source Code Yes Code is available at: https://github.com/ Vivi Xiang/Social CVAE.
Open Datasets Yes To evaluate the effectiveness of our method, we conduct extensive experiments on two widely used datasets in pedestrian trajectory prediction tasks: ETH-UCY dataset (Pellegrini et al. 2009; Lerner, Chrysanthou, and Lischinski 2007) and Stanford Drone Dataset (SDD) (Robicquet et al. 2016).
Dataset Splits Yes We follow the leave-one-out strategy (Mangalam et al. 2021) for training and evaluation. SDD contains pedestrians trajectories in 20 scenes. For SDD, we follow the data segmentation as (Yue, Manocha, and Wang 2022) for training and evaluation. Following the common practice (Mangalam et al. 2021; Yue, Manocha, and Wang 2022), the raw trajectories are segmented into 8-second trajectory segments with time step t = 0.4s, we train the model to predict the future 4.8s (12 frames) based on the observed 3.2s (8 frames).
Hardware Specification Yes Our model was implemented in PyTorch on a desktop computer running Ubuntu 20.04 containing an Intel Core TM i7 CPU and an NVIDIA GTX 3090 GPU.
Software Dependencies No The paper states 'Our model was implemented in PyTorch' but does not provide a specific version number for PyTorch or any other software libraries/dependencies.
Experiment Setup Yes The model is trained end-to-end with an Adam optimizer with a learning rate 0.0001. We trained the ETH-UCY for 100 epochs and SDD for 150 epochs.