Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels

Authors: Zhizheng Liu, Joe Lin, Wayne Wu, Bolei Zhou

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that Ped Gen outperforms existing baseline methods for pedestrian movement generation by learning from noisy labels and incorporating the context factors. In addition, Ped Gen achieves zero-shot generalization in both real-world and simulated environments. Experiment results on the City Walkers validation set, the real-world Waymo open dataset (Sun et al., 2020) and CARLA simulator (Dosovitskiy et al., 2017) show that Ped Gen can predict more realistic and accurate future pedestrian movements than existing human motion generation methods and achieve better zero-shot generalization by generating high-quality context-aware movements. Additional experiments and ablation studies demonstrate the effectiveness of Ped Gen in addressing noisy labels and incorporating the key context factors.
Researcher Affiliation Academia Zhizheng Liu, Joe Lin, Wayne Wu, Bolei Zhou Department of Computer Science, University of California, Los Angeles
Pseudocode No The paper describes the methods in prose and mathematical formulations within sections like "4 GENERATING CONTEXT-AWARE PEDESTRIAN MOVEMENTS FROM NOISY LABELS" and its subsections. It does not contain any explicitly labeled "Pseudocode" or "Algorithm" blocks or figures.
Open Source Code Yes The code, model, and data are available at https://genforce.github.io/Ped Gen/.
Open Datasets Yes In this work, we propose learning diverse pedestrian movements from web videos. We first curate a large-scale dataset called City Walkers that captures diverse real-world pedestrian movements in urban scenes. Then, based on City Walkers, we propose a generative model called Ped Gen for diverse pedestrian movement generation. The code, model, and data are available at https://genforce.github.io/Ped Gen/. Experiment results on the City Walkers validation set, the real-world Waymo open dataset (Sun et al., 2020) and CARLA simulator (Dosovitskiy et al., 2017) show that Ped Gen can predict more realistic and accurate future pedestrian movements...
Dataset Splits Yes The training set has 104,192 samples, including 53,405 partial labels that have at least 30 frames of annotation. The validation set has 13,039 samples and only contains complete labels.
Hardware Specification No The paper discusses experiments and model training, but does not specify any particular hardware (e.g., GPU models, CPU types, memory) used for these processes.
Software Dependencies No The paper mentions several models and tools, such as "diffusion models (Ho et al., 2020)", "Adam (Kingma & Ba, 2014) optimizer", "WHAM (Shin et al., 2023)", "Zoe Depth (Bhat et al., 2023)", "SegFormer (Xie et al., 2021)", and "YOLO (Jocher et al., 2023)". However, it does not provide specific version numbers for any of these software components.
Experiment Setup No The paper provides details on data sampling for training: "For each pedestrian movement trajectory, we sample the initial timestep at an interval of 30 frames and keep at most the future 60 frames (2 seconds) as the ground truth movement." It also mentions dataset sizes for training and validation. However, it lacks specific hyperparameter values such as learning rate, batch size, or number of training epochs within the main text provided.