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].

RedMotion: Motion Prediction via Redundancy Reduction

Authors: Royden Wagner, Omer Sahin Tas, Marvin Klemp, Carlos Fernandez, Christoph Stiller

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments reveal that our representation learning approach outperforms Pre Tra M, Traj-MAE, and Graph DINO in a semi-supervised setting. Moreover, Red Motion achieves competitive results compared to HPTR or MTR++ in the Waymo Motion Prediction Challenge. Our open-source implementation is available at: https://github.com/kit-mrt/future-motion
Researcher Affiliation Academia 1Karlsruhe Institute of Technology 2FZI Research Center for Information Technology
Pseudocode No The paper describes the method in detail with figures and text, but it does not include any specific pseudocode blocks or algorithms.
Open Source Code Yes Our open-source implementation is available at: https://github.com/kit-mrt/future-motion
Open Datasets Yes We use the official training and validation splits of the Waymo Open Motion dataset (Ettinger et al., 2021) version 1.0 and the Argoverse 2 Forecasting dataset (Wilson et al., 2021) as training and validation data.
Dataset Splits Yes We use the official training and validation splits of the Waymo Open Motion dataset (Ettinger et al., 2021) version 1.0 and the Argoverse 2 Forecasting dataset (Wilson et al., 2021) as training and validation data. Since pre-training is particularly useful when little annotated data is available, we use 100% of the training data for pre-training and fine-tune on only 12.5%, following common practice in self-supervised learning (Balestriero et al., 2023).
Hardware Specification Yes We pre-train and fine-tune all configurations for 4 hours and 8 hours using data-parallel training on 4 A100 GPUs.
Software Dependencies No The paper mentions using Adam W as an optimizer and PyTorch for implementation, but it does not specify version numbers for these software components or any other libraries.
Experiment Setup Yes For pre-training and fine-tuning, we use Adam W (Loshchilov & Hutter, 2019) as the optimizer. The initial learning rate is set to 10^-4 and reduced to 10^-6 using a cosine annealing learning rate scheduler (Loshchilov & Hutter, 2016). We pre-train and fine-tune all configurations for 4 hours and 8 hours using data-parallel training on 4 A100 GPUs. Following Konev et al. (2022), we minimize the negative multivariate log-likelihood loss for fine-tuning on motion prediction. For Traj-MAE pre-training, we mask 60% of the road environment tokens and train to reconstruct them. Our model is computed considering 6 trajectory proposals per agent. We use an attention window size of 16 tokens.