Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Neuralized Markov Random Field for Interaction-Aware Stochastic Human Trajectory Prediction
Authors: Zilin Fang, David Hsu, Gim H Lee
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our proposed method achieves stateof-the-art performance on ADE/FDE metrics across two dataset categories: overhead datasets ETH/UCY, SDD, and NBA, and ego-centric JRDB. Furthermore, our approach allows for real-time stochastic inference in bustling environments, making it well-suited for a 30FPS video setting. We open-source our codes at: https://github.com/Ada Comp NUS/NMRF Trajectory Prediction.git. |
| Researcher Affiliation | Academia | Zilin Fang1, David Hsu1,2, Gim Hee Lee1 1School of Computing, National University of Singapore (NUS) 2Smart Systems Institute, NUS |
| Pseudocode | No | The paper describes the network architecture and its components (e.g., CVAEs, History Encoder, Update Decoder, Configuration Encoder, Dynamics Decoder) but does not present explicit pseudocode or algorithm blocks for the overall methodology. |
| Open Source Code | Yes | Our proposed method achieves stateof-the-art performance on ADE/FDE metrics across two dataset categories: overhead datasets ETH/UCY, SDD, and NBA, and ego-centric JRDB. Furthermore, our approach allows for real-time stochastic inference in bustling environments, making it well-suited for a 30FPS video setting. We open-source our codes at: https://github.com/Ada Comp NUS/NMRF Trajectory Prediction.git. |
| Open Datasets | Yes | We evaluate our methods on four trajectory prediction datasets: ETH/UCY (Pellegrini et al., 2009; Lerner et al., 2007), Stanford Drone Dataset (SDD) (Robicquet et al., 2016), NBA Sport VU Dataset (NBA), and the Jack Rabbot Dataset and Benchmark (JRDB) (Martin-Martin et al., 2021), covering interaction-rich indoor and outdoor scenarios. |
| Dataset Splits | Yes | ETH/UCY dataset contains five subsets: ETH, HOTEL, UNIV, ZARA1, and ZARA2. We follow the leave-one-out approach from (Gupta et al., 2018) with four subsets for training-validation and the remaining subset for testing, predicting the future 12 frames (4.8s) using 8 frames observations (3.2s). ... SDD ... Train-test splits are the same as the baseline Social-VAE (Xu et al., 2022c). ... JRDB ... We follow the train-validation-test split applied in Social-Transmotion (Saadatnejad et al., 2023) for deterministic prediction. For stochastic situations, splits are in accordance with the official JRDB detection and tracking challenge. |
| Hardware Specification | Yes | We set sample number N to 20 in both training and testing stages for stochastic prediction, and train the entire model with Adam W optimizer and Step LR scheduler on one Quadro RTX 8000 GPU. |
| Software Dependencies | No | The paper mentions using 'Adam W optimizer' and 'Step LR scheduler' as well as 'GRU' and 'multilayer perceptions' as components, but does not specify software library versions (e.g., PyTorch version, Python version, specific deep learning framework version). |
| Experiment Setup | Yes | In the training phase, we set hyperparameters α, β and λ in loss terms as 1, 1 10 3 and 1 10 4. ... The latent vectors Z1 and Z2 have 16 dimensions and the embedding dimension of the configuration encoder is 32. ... Table 9: Hyperparameters for different datasets. Dataset Stride Batch Size Learning Rate Step Size Gamma Epoch (CVAE) Epoch (sampler) ETH 3 64 2 10 4 16 0.5 200 60 HOTEL 3 50 2 10 4 16 0.5 200 60 UNIV 3 32 2 10 4 32 0.9 200 60 ZARA1 3 32 2 10 4 32 0.9 200 60 ZARA2 3 32 2 10 4 32 0.9 200 60 SDD 3 16 8 10 4 32 0.9 500 200 NBA 10 32 1 10 4 32 0.6 150 50 JRDB 3 16 1 10 4 32 0.6 200 50 |