Probabilistic Tracklet Scoring and Inpainting for Multiple Object Tracking

ArTIST

Probabilistic Tracklet Scoring and Inpainting for Multiple Object Tracking (CVPR 2021)

Pytorch implementation of the ArTIST motion model. In this repo, there are

Training script for the Moving Agent network
Training script for the ArTIST motion model
Demo script for Inferring the likelihood of current observations (detections)
Demo script for Inpainting the missing observation/detections

Demo 1: Likelihood estimation of observation

Run:

python3 demo_scoring.py

This will generate the output in the temp/ar/log_p directory, look like this:

likelihood

This demo gets as input a pretrained model of the Moving Agent Network (MA-Net), a pretrained model of ArTIST, the centroids (obtain centroids via the script in the utils), a demo test sample index and the number of clusters.

The model then evaluates the log-likelihood (lower the better) of all detections as the continuation of the observed sequence.

Demo 2: Sequence inpainting

Run:

python3 demo_inpainting.py

This will generate the multiple plauusible continuations of an observed motion, stored in the temp/ar/inpainting directory. One example looks like this:

inpainting_25

For each generated future sequence, it computes the IoU between the last generated bounding box and the last groundtruth bounding box, as well as the mean IoU for the entire generated sequence and the groundtruth sequence.

Utilities

In this repo, there are a number of scripts to generate the required data to train/evaluate ArTIST.

prepare_data: Given the annotations of a dataset (e.g., MOT17), it extracts the motion sequences as well as the IDs of the social tracklets living the life span of the corresponding sequence, and stores it as a dictionary. If there are multiple tracking datasets that you wish to combine, you can use the merge_datasets() function inside this script.
clustering: Given the output dictionary of prepare_data script, this script performs the K-Means clustering and stores the centroids which are then used in the ArTIST model.
dataloader_ae and dataloader_ar: Given the post-processes version of the dataset dictionary (which can be done by running the post_process script), these two scripts define the dataloaders for training the MA-Net and ArTIST. Note that the dataloader of ArTIST uses the MA-Net to compute the social information. This can also be done jointly in an end-to-end fashion, which we observed almost no difference.
create_demo_test_subset: In order to run the demo scripts, you need to run this script. However, the demo test subset has been produced and stored in data/demo_test_subset.npy.

Data

You can download the required data from the Release and put it in data/ directory.

Citation

If you find this work useful in your own research, please consider citing:

@inproceedings{saleh2021probabilistic,
author={Saleh, Fatemeh and Aliakbarian, Sadegh and Rezatofighi, Hamid and Salzmann, Mathieu and Gould, Stephen},
title = {Probabilistic Tracklet Scoring and Inpainting for Multiple Object Tracking},
booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
year = {2021}
}