Semi-Supervised Generative Models for Multiagent Trajectories
Authors: Dennis Fassmeyer, Pascal Fassmeyer, Ulf Brefeld
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, our model not only outperforms various state-of-the-art baselines in trajectory forecasting, but also learns to effectively leverage unsupervised multi-agent sequences for classification tasks on interactive real-world sports datasets. 4 Empirical Evaluation For evaluation, we focus on team sports as the coordination of players renders these tasks more difficult than other domains. Hence, we experiment on STATS Sport VU NBA4 for comparison, and tracking data from soccer games of the german national team. |
| Researcher Affiliation | Academia | Dennis Fassmeyer Leuphana University of Lüneburg dennis.fassmeyer@leuphana.de Pascal Fassmeyer Leuphana University of Lüneburg pascal.fassmeyer@leuphana.de Ulf Brefeld Luephana University of Lüneburg brefeld@leuphana.de |
| Pseudocode | No | The paper describes the model conceptually and mathematically but does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/fassmeyer/MAT_Neur IPS22. (Footnote 5 in Section 4) |
| Open Datasets | Yes | we experiment on STATS Sport VU NBA4 for comparison, and tracking data from soccer games of the german national team. As detailed in Section 3.2, we use agent velocities as input to all models and assume linear motion between consecutive observations. For NBA, we adopt the experimental setup and processing strategy from [38]5. The STATS Sport VU NBA data comprises tracking positions of offensive plays from the 2016 NBA regular season covering more than 1200 different games where game segment are given by sequences of length 50 and contain two-dimensional positions of all agents (10 players and ball) sampled at 5 frames per second. The data is split into 60% training, 20% validation, and 20% test sets. All data is translated so that the origin of the underlying coordinate system is mapped onto the top-left corner. The soccer data consists of 12 matches where positions of players and ball are sampled at 25 frames per second. The tracking data is accompanied by manual event annotations that we will also make use of in the remainder. Models are trained on eight matches, the remaining four are distributed evenly into validation and test data (two matches each). ... 4https://github.com/linouk23/NBA-Player-Movements |
| Dataset Splits | Yes | The data is split into 60% training, 20% validation, and 20% test sets. (NBA data in Section 4) Models are trained on eight matches, the remaining four are distributed evenly into validation and test data (two matches each). (Soccer data in Section 4) |
| Hardware Specification | No | Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix (Section 1 of the checklist at the end of the paper). |
| Software Dependencies | No | The paper describes the models and architectures used (e.g., RNN, GNN, LSTM, Info-GAN) but does not provide specific version numbers for software dependencies or libraries. |
| Experiment Setup | Yes | The data is split into 60% training, 20% validation, and 20% test sets. All data is translated so that the origin of the underlying coordinate system is mapped onto the top-left corner. (Section 4, NBA data). Following related work [38, 60], we obtain 90 different areas/class labels. (Section 4.2, S-MAT setup). We maximize Eqn (4) using 3 agent types and refer to the unsupervised variant as U-MAT. (Section 4.2, U-MAT setup). Error for NBA in meters for a prediction interval of 10 timesteps with an observation period of 40 timesteps. (Table 1 caption). We focus on labels Y = {pass, other ball action, shot, none}, where class none denotes the absence of all other labels in a frame and use 20% annotated data. Every label is propagated to the previous five and the subsequent 30 frames. We generate a balanced training set where half of the segments carry label none. (Section 4.4, Soccer data setup). |