Learning Cooperative Trajectory Representations for Motion Forecasting

Authors: Hongzhi Ruan, Haibao Yu, Wenxian Yang, Siqi Fan, Zaiqing Nie

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

Reproducibility Variable Result LLM Response
Research Type Experimental V2X-Graph is evaluated on V2X-Seq in vehicle-to-infrastructure (V2I) scenarios. To further evaluate on vehicle-to-everything (V2X) scenario, we construct the first real-world V2X motion forecasting dataset V2X-Traj, which contains multiple autonomous vehicles and infrastructure in every scenario. Experimental results on both V2X-Seq and V2X-Traj show the advantage of our method.
Researcher Affiliation Academia Hongzhi Ruan 1,2 Haibao Yu 1,3 Wenxian Yang 1 Siqi Fan 1 Zaiqing Nie 1 1 Institute for AI Industry Research (AIR), Tsinghua University 2 University of Chinese Academy of Science 3 The University of Hong Kong
Pseudocode Yes Algorithm 1: Pseudo Labels Generator Input: Ego-view trajectories Tego, Other-view trajectories Tother Output: Cross-views trajectories matching pesudo labels A
Open Source Code Yes Find the project at https://github.com/AIR-THU/V2X-Graph.
Open Datasets Yes (1) V2X-Seq [51]. A public large-scale and real-world V2I dataset. (2) V2X-Traj (Ours). To study the effectiveness of V2X-Graph in V2V and broader V2X scenarios, especially its ability to handle more than two views of trajectories, including both V2I and V2V cooperation, we construct the first real-world and public V2X cooperative motion forecasting dataset, termed V2X-Traj.
Dataset Splits Yes V2X-Traj dataset contains a total of 10,102 scenarios, which are randomly split into the training, validation, and test set, consisting of 6,062, 2,020, and 2,020 scenarios, respectively.
Hardware Specification Yes The model is trained for 64 epochs with batch size of 64 on a server with 8 NVIDIA RTX 4090s. Then we conduct the inference experiment on single NVIDIA GTX 4090 and compare the inference cost.
Software Dependencies No The paper mentions using 'Adam W optimizer' and 'Transformer' modules, and re-implementing 'Hi VT', 'Dense TNT', and 'HDGT', but it does not specify version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or other key dependencies.
Experiment Setup Yes For training, the initial learning rate is set to 1 × 10^−3 and is scheduled according to cosine annealing [27]. The Adam W optimizer [28] is adopted with a weight decay of 1 × 10^−4. The model is trained for 64 epochs with batch size of 64 on a server with 8 NVIDIA RTX 4090s. The dimensions of the hidden feature is set as 128, and the number of heads in all multi-head attention blocks is 16.