TrEP: Transformer-Based Evidential Prediction for Pedestrian Intention with Uncertainty
Authors: Zhengming Zhang, Renran Tian, Zhengming Ding
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on three popular pedestrian intent benchmarks have verified the effectiveness of our proposed model over the state-of-the-art. The algorithm performance can be further boosted by controlling the uncertainty level. We systematically compare human disagreements with AI uncertainty to further evaluate AI performance in confusing scenes. |
| Researcher Affiliation | Academia | Zhengming Zhang1, Renran Tian2, Zhengming Ding3 1 School of Industrial Engineering, Purdue University, West Lafayette, Indiana, USA 2 Department of Computer Information Technology, Indiana University Purdue University Indianapolis, Indiana, USA 3 Department of Computer Science, Tulane University, New Orleans, Louisiana, USA |
| Pseudocode | No | The paper describes its models (Base Model, Uncertainty-Aware Evidential Learning) using text and mathematical equations, and provides an overview diagram (Figure 1), but it does not include a dedicated pseudocode block or algorithm listing. |
| Open Source Code | Yes | The code is released at https://github.com/zzmonlyyou/Tr EP.git. |
| Open Datasets | Yes | In our experiments, three intention/action prediction benchmarks are explored, which are JAAD (Rasouli, Kotseruba, and Tsotsos 2017), PIE (Rasouli et al. 2019), and PSI (Chen et al. 2021). To our best knowledge, those three benchmarks are the most representative datasets regarding the intent prediction task. |
| Dataset Splits | No | The paper specifies the number of training sequences for each dataset (e.g., 'PIE has 3,980 training sequences', 'JAAD has 3,955 training sequences', 'PSI has 6,262 training sequences'). However, it does not explicitly state percentages or absolute counts for validation splits or general train/validation/test splits. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions the use of 'Adam optimizer' but does not specify its version or any other software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow versions). |
| Experiment Setup | Yes | Due to the different annotations and feature engineering for the model on each dataset, the input dimensions (b t fd) are slightly different (where b refers to batch size (we set b = 64), t and fd refers to the size of time span and feature dimension). We projected the input features dimensions fd to 8 dimensions in the first linear layer. The fully connected layers in the transformer projected the 8 dimensions to 16. There is one layer of multi-head attention (2 heads) for PIE and PSI and two layers for JAAD. The dropout rates are set to 0.1. All the models are trained by Adam optimizer with a learning rate of 5e-3 for 2,000 epochs. |