Generative Causal Representation Learning for Out-of-Distribution Motion Forecasting
Authors: Shayan Shirahmad Gale Bagi, Zahra Gharaee, Oliver Schulte, Mark Crowley
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on synthetic and real-world motion forecasting datasets show the robustness and effectiveness of our proposed method for knowledge transfer under zero-shot and low-shot settings by substantially outperforming the prior motion forecasting models on out-of-distribution prediction. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada 2Department of Systems Design Engineering, University of Waterloo, Waterloo, Canada 3School of Computing Science, Simon Fraser University, Burnaby, Canada. |
| Pseudocode | No | The paper does not contain structured pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/sshirahmad/GCRL. |
| Open Datasets | Yes | ETH-UCY dataset This dataset contains the trajectory of 1536 detected pedestrians captured in five different environments {hotel, eth, univ, zara1, zara2}. All trajectories in the dataset are sampled every 0.4 seconds. Following the experimental settings of (Liu et al., 2022; Chen et al., 2021; Huang et al., 2019), we also use a leave-one-out approach for training and evaluating our model so to predict the future 4.8 seconds (12 frames), we utilize the previously observed 3.2 seconds (8 frames). |
| Dataset Splits | Yes | Each domain contains 10,000 trajectories for training, 3,000 trajectories for validation, and 5,000 trajectories for testing. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running its experiments, such as exact GPU/CPU models, processor types, or memory amounts. |
| Software Dependencies | No | The paper does not provide specific software dependency details with version numbers (e.g., Python 3.x, PyTorch x.x, CUDA x.x) needed to replicate the experiment environment. |
| Experiment Setup | Yes | The list of all hyperparameters used by our model and their corresponding settings applied when conducting our experiments are represented in Tables 5 and 6. Our model is trained for 300 epochs in experiments conducted with the ETH-UCY dataset. We trained our method GCRL in the experiments with synthetic dataset for 250 epochs. We fine-tuned the trained models for the domain adaptation task for 100 epochs. The batch size is 64. |