Diverse Trajectory Forecasting with Determinantal Point Processes
Authors: Ye Yuan, Kris M. Kitani
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the diversity of the trajectories produced by our approach on both low-dimensional 2D trajectory data and high-dimensional human motion data. |
| Researcher Affiliation | Academia | Ye Yuan, Kris M. Kitani Robotics Institute Carnegie Mellon University {yyuan2,kkitani}@cs.cmu.edu |
| Pseudocode | Yes | Algorithm 1 Training the diversity sampling function (DSF) Sγ(ψ) |
| Open Source Code | No | The paper does not provide a link to the source code for the described methodology or state that it is open-source. A video link is provided for qualitative results, but not for code. |
| Open Datasets | Yes | We use motion capture to obtain 10 motion sequences including different types of motions such as walking, turning, jogging, bending, and crouching. Each sequence is about 1 minute long and each pose consists of 59 joint angles. We use past 3 poses (0.1s) to forecast next 30 poses (1s). There are around 9400 training examples and 2000 test examples where we use different sequences for training and testing. More implementation details can be found in Appendix B. For additional experiments, it states: We also perform additional experiments on a large human motion dataset (3.6 million frames), Human3.6M (Ionescu et al., 2013). |
| Dataset Splits | No | The paper mentions 'training examples' and 'test examples' and specifies train/test splits (e.g., '1100 training examples and 1000 test examples' or 'We train on five subjects (S1, S5, S6, S7, S8), and test on two subjects (S9 and S11)'). However, it does not explicitly mention a separate 'validation' dataset split. |
| Hardware Specification | No | The paper describes network architectures (e.g., CNN, MLP, Bi-LSTMs) and training procedures (e.g., optimizers, learning rates, batch sizes), but it does not specify any hardware details like GPU models, CPU models, or cloud computing instances used for the experiments. |
| Software Dependencies | No | The paper mentions using 'Adam (Kingma and Ba, 2014)' as an optimizer, but it does not specify software dependencies like programming language versions, specific deep learning framework versions (e.g., PyTorch, TensorFlow), or other library versions. |
| Experiment Setup | Yes | The weighting factor β for the KL term is set to 0.1 for synthetic data and 1e-4 for human motion. We use Adam (Kingma and Ba, 2014) to jointly optimize the encoder and decoder. The learning rate is set to 1e-4 and we use a mini batch size of 32 for synthetic data. We optimize the model for 500 epochs for synthetic data and 100 epochs for human motion. The scale factor k for the similarity matrix S is set to 1 for synthetic data and 1e-2 for human motions. For both synthetic data and human motions, we use Adam with learning rate 1e-4 to optimize the DSF for 20 epochs. |