Open-World-Tracking

Official code for "Opening up Open-World Tracking" (CVPR 2022 Oral)

The repository contains the code for the experiments we conducted in the paper. We also provide the baseline trackers.

Opening up Open-World Tracking
Yang Liu*, Idil Esen Zulfikar*, Jonathon Luiten*, Achal Dave*, Deva Ramanan, Bastian Leibe, Aljoša Ošep, Laura Leal-Taixé
*Equal contribution
CVPR 2022 Oral

1. Dependencies

Python >= 3.7
PyTorch==1.9.1
detectron2 installation, see this
tested on CUDA=11.1

You can also check requirements.txt for more details.

2. Data and Models

Data Download

We use TAO as our base dataset, so please follow the instructions here to download the raw data and the labels.

Known and Unknown Split

As stated in the paper, we use COCO-classes as known, you can find the mapping between coco_id to tao_id in datasets/coco_id2tao_id.json. Keeping the default folder structure is fine.

To get the known/unknown ids in TAO:

import json

json_path = "./datasets/coco_id2tao_id.json"
with open(json_path, 'r') as f:
    coco2tao_map = json.load(f)
knowns = {v for _, v in coco2tao_map.items()}

json_path = "./datasets/distractor_classes.json"
with open(json_path, 'r') as f:
    coco2distractor_map = json.load(f)

distractors = []
for _, v in coco2distractor_map.items():
    distractors += v
distractors_ids = set(distractors)

all_ids = set([i for i in range(1, 2000)])  # 2000 is larger than the max category id in TAO-OW.
unknowns = all_ids.difference(knowns.union(distractors))

Model

We adapt detectron2 to generate per-frame proposals. Please visit https://github.com/YangLiu14/detectron2-OWT

The detailed instructions/model config/model weights are in README in the repo mentioned above.

3. Running the Tracker

The detailed instruction is in trackers/README.md. Give the correct information in config.yaml and run:

python nonoverlap_track.py --tracker online-ghost --config ./configs/example_tracker.yaml

The tracking results will be written as .txt files.

For evaluation, you need to aggregate the results in all the txt files in one single .json file.

You can use tools/track_results_conversion.py to do that.

python tools/track_result_conversion.py \
  --results_dir /results/my_tracker/ \
  --gt_path /data/TAO/annotations/validations.json \
  --outname tracking \
  --split val

After running this script, you will get /results/my_tracker/_COCO/tracking.json

4. Evaluation

For instructions for evaluation, please visit link

download/clone the TrackEval

git clone https://github.com/JonathonLuiten/TrackEval.git

Copy the GT annotation from /data/TAO/annotations/validation.json to a new folder (say /data/TAO/trackeval/validations.json). Make sure this folder only contains one annotation file.
Suppose your tracking results (COCO format, see section 3) are stored in /results/my_tracker/_COCO/tracking.json

cd scripts/

python run_tao_ow.py \
  --GT_FOLDER /data/TAO/trackeval/ \
  --TRACKERS_FOLDER results \
  --TRACKERS_TO_EVAL my_trackers \
  --SPLIT_TO_EVAL val \
  --TRACKER_SUB_FOLDER _COCO \
  --TRACKER_DISPLAY_NAMES MY_TRACKER \
  --MAX_DETECTIONS 1000 \
  --SUBSET unknown

Citation

@inproceedings{liu2022opening,
  title={Opening up Open-World Tracking},
  author={Liu, Yang and Zulfikar, Idil Esen and Luiten, Jonathon and Dave, Achal and Ramanan, Deva and Leibe, Bastian and O{\v{s}}ep, Aljo{\v{s}}a and Leal-Taix{\'e}, Laura},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

YangLiu14 / Open-World-Tracking