TrackingNet-devkit

Development kit for the dataset TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild.

Compete in our benchmark by submitting your result on our evaluation server.

For more details, please refer to our paper.

@InProceedings{Muller_2018_ECCV,
author = {Muller, Matthias and Bibi, Adel and Giancola, Silvio and Alsubaihi, Salman and Ghanem, Bernard},
title = {TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}

[Update from Jan 24th, 2021]: pip package

We have developed a pip package for automatic download and utils function on TrackingNet. You can install it simply with:

pip install TrackingNet

More details on https://pypi.org/project/TrackingNet/.

[Update from May 2nd, 2020]: Backup ownCloud solution

We have a new ownCloud solution to download the different splits for training/testing:

https://exrcsdrive.kaust.edu.sa/exrcsdrive/index.php/s/MAaiTPdOwiPDNlp

The password is: TrackingNet.

For maximum bandwidth performances, use the option "Add to your ownCloud" with any ownCloud account.

[Update from April 22nd, 2020]: Bittorent solution

We are currently experimenting a bittorrent solution to share TrackingNet among the tracking community. Multiple torrents are available on https://academictorrents.com/collection/trackingnet.

It may be very slow at the beginning, but it will improve once more people will require the a copy. Here are some guidelines:

• I hope everyone will play fair and seed its torrent as we are currently doing.
• Feel free to share the torrent between colleagues, seedbox or mirroring server.
• The google drive backup links [link1] [link2] are still available, but capped in daily bandwidth.
• If you already have downloaded a chunk from the backup links, please put it in the output folder of your torrent: it won't download it again and will save you precious download time.

PS: In case your university/institution blocks your torrents, you can rent a seedbox that will download it for you. It's legal, cheap (<15$ for a full month) and have very fast bandwidth (up to 125MB/sec - Full TrackingNet downloaded in 2 hours). I recommend seedbox.io (currently used to seed at up to 125MB/sec), which provides a direct FTP connection to access TrackingNet once it's downloaded. Fancier seedboxes can support Google Drive, Baidu Pan, Dropbox, ownCloud, AWS S3, etc...

[Update from February 25th, 2020]: Back up links

To anyone who still has issue downloading TrackingNet, we are currently trying to find more reliable solutions. For now, we have created back up links to download full chunks of training (and the testing chunk). It's still hosted on Google Drive, but will be easier to spread around the community using alternative sharing platforms (e.g. Baidu, Dropbox, good old HDD,...).

Here are two back up links: [link1] [link2]

Now, it appears that Google Drive is limiting the download if you are not signed in with you gmail account. If you have any issue downloading it, please make sure you are signed in google drive with you gmail account. We will track the situation in the next days.

Structure of the dataset

There are 12 chunks of 2511 sequences for the training and 1 chunk of 511 sequences for the testing.

Each chunk have subfolders for the zipped sequence (zips), the unzipped frame (frames) and eventually the annotation (anno).

The structure of the dataset is the following:

TrackingNet
 - Test / Train_X (with X from 0 to 11)
   - zips
   - frames
   - anno (Test: annotation only for 1st frame)

Create the environment

Tested on Ubuntu 16.04 LTS

• Create the environment:

conda env create -f environment.yml

or (preferred for other platforms)

conda create -n TrackingNet python=3 requests pandas tqdm numpy

• Activate the environment:

source activate TrackingNet (activate TrackingNet for windows platforms)

Download the dataset

You can download the whole dataset by running:

python download_TrackingNet.py --trackingnet_dir <trackingnet_dir>

Optional parameters:

• --trackingnet_dir: path where to download the TrackingNet dataset
• --data select the data to download (sequences: --data zips / annotations: --data anno)
• --chunk select the chunk to download (testing set: --chunk Test / training set: --chunk Train / selected chunks: --chunk 0,2,4,11)

Please look at python download_TrackingNet.py --help for more details on the optional parameters.

Disclaimer

In case an error such as Permission denied: https://drive.google.com/uc?id=<ID>, Maybe you need to change permission over 'Anyone with the link'? occurs, please check your internet connection and run again the script. The script will not overwrite the previous sequences of videos if are already completely downloaded.

Note that Google Drive limits the download bandwidth to ~10TB/day. To ensure a good share between all users, avoid downloading the dataset several times and prefer sharing it with your colleagues using an old-fashion HDD.

Further downloading options:

• Google Drive (backup links): [link1] [link2]
• Bittorrent: https://academictorrents.com/details/1faf1b53cc0099d2206f02be42b5688952c3c6b3

Unzip the frames

To extract all the zipped sequences for the complete dataset:

python extract_frame.py --trackingnet_dir <trackingnet_dir>

Optional parameters:

• --trackingnet_dir: path where to download the TrackingNet dataset
• --chunk: select the chunk to download (testing set: --chunk Test / training set: --chunk Train / selected chunks: --chunk 0,2,4,11)

Disclaimer

In this step, make sure you don't have any error message. You can run this script several times to make sure all the files are properly extracted. By default, the unzipping script will not overwrite the frames that are properly extracted.

If any zip file is currupted, a error message will appear Error: the zip file [zip_file_name] is corrupted. In thas case, remove the corrupted zip file manually and run the download script again. By default, the download script will not overwrite the zip files already downloaded.

(Optional) Generate Frames with the annotation boundingboxes

This part requires opencv: conda install -c menpo opencv

To generate the BB in the frames for the complete dataset:

python generate_BB_frames.py --output_dir <trackingnet_dir>

Optional parameters:

• --output_dir: path where to generate the images with boundingboxes
• --chunk select the chunk to download (testing set: --chunk Test / training set: --chunk Train / selected chunks: --chunk 0,2,4,11)

Evaluate the results of a tracker with a given ground truth

If you plan to submit results on our evaluation server, you may want to validate your results first.

The evaluation code we are using is available on metrics.py, whhich can be used as following:

python metrics.py --GT_zip <GT.zip> --subm_zip <subm.zip>

A dummy example of file is provided here:

python metrics.py --GT_zip dummy_GT.zip --subm_zip dummy_subm.zip