This repo provides a short .ipynb tutorial for loading and visualizing the CARLA data from the PRECOG paper.
Paper pdf link: https://arxiv.org/pdf/1905.01296.pdf
Project page link: https://sites.google.com/view/precog
Visualizations on nuScenes data:
Link to PRECOG implementation code: https://github.com/nrhine1/precog
Link to Deep Imitative Models implementation code: https://github.com/nrhine1/deep_imitative_models
Here's a link to the dataset we used for training and evaluating our approach on CARLA Town01 (see Appendix for experiment details) https://drive.google.com/drive/folders/1arYDYuG4SDrQkc4ynzVOG5ObEOVvq9iX
Here's a link to the dataset we used for training and evaluating our approach on CARLA Town01 (see Appendix for experiment details) https://drive.google.com/drive/folders/1s-g5hWtDLrcQPjc6Xra5Up3BFF3jrJle
The overhead_features is a very simple featurization of the 3D LIDAR point cloud. I did not put much effort into tuning this representation. Nonetheless, here's the format:
Channel 0: Histogram of all points in each cell (for all heights)
Channel 1: Histogram of points above ground z-threshold
Channel 2: Histogram of points below ground z-threshold
Channel 3: Histogram of points above a higher z-threshold
The z-thresholds are listed here: https://github.com/nrhine1/deep_imitative_models/blob/00419760f50ca0800be7a95716d75a5cf421867f/dim/env/preprocess/carla_preprocess.py#L611-L613
They are -4.5, and -2.0 in the coordinate frame of the LIDAR sensor.