KITTI Dataset for trajectory prediction

This repository contains the Kitti dataset for trajectory prediction used in "MANTRA: Memory Augmented Networks for Multiple Trajectory Prediction" published at CVPR2020 (arxiv)

KITTI dataset

To obtain samples, we collect 6 seconds chunks (2 seconds for past and 4 for future) in a sliding window fashion from all trajectories in the dataset, including the egovehicle.

For the training set, we use the following videos (KITTI enumeration): 5, 9, 11, 13, 14, 17, 27, 28, 48, 51, 56, 57, 59, 60, 84, 91. For the test set, the remaining videos: 1, 2, 15, 18, 29, 32, 52, 70.

We obtain 8613 top-view trajectories for training and 2907 for testing. Further details are in the paper (section 4.1).

FILE DESCRIPTION

The files dataset_KITTI_train.json and dataset_KITTI_test.json contain respectively the train and test set. The dataset_pytorch.py file is a script to create a pytorch-style dataset with class and methods to visualize the single examples.

The maps folder contains whole top-view context of a specific video. For each example, we crop an area relative (180x180 meters) to the present position of the agent to predict.

USAGE

To create a dataset and relative dataloader to train your own models:

from torch.utils.data import DataLoader
import dataset_pytorch
from tqdm import tqdm

data_train   = dataset_pytorch.TrackDataset('dataset_kitti_test.json')
train_loader = DataLoader(data_train, batch_size=32, num_workers=1, shuffle=True)
data_test    = dataset_pytorch.TrackDataset('dataset_kitti_test.json')
test_loader  = DataLoader(data_test, batch_size=32, num_workers=1, shuffle=False)

for step, (index, past, future, scene_one_hot, video, class, num_vehicles, step, scene) in enumerate(tqdm(train_loader)):
    #code to call own model
              

scene_one_hot ([dim_batch,180,180,4] dimension) is used for training while scene ([dim_batch,180,180,1] dimension) is used for qualitative visualization.

AUTHORS AND CONTACTS

• Francesco Marchetti (MICC - Università degli Studi di Firenze)
• Federico Becattini (MICC - Università degli Studi di Firenze)
• Lorenzo Seidenari (MICC - Università degli Studi di Firenze)
• Alberto Del Bimbo (MICC - Università degli Studi di Firenze)

For questions and explanations, you can contact by e-mail to francesco.marchetti@unifi.it

If you use this code, please cite the paper:

@inproceedings{marchetti2020mantra,
  title={Mantra: Memory augmented networks for multiple trajectory prediction},
  author={Marchetti, Francesco and Becattini, Federico and Seidenari, Lorenzo and Bimbo, Alberto Del},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7143--7152},
  year={2020}
}