Tracking has matured to a point where the constraints enforced in past VOT challenges can be relaxed and general object tracking should be considered in a broader context. The new Visual Object Tracking and Segmentation challenge VOTS2023 thus no longer distincts between single- and multi-target tracking nor between short- and long-term tracking. We propose a single challenge that requires tracking one or more targets simultaneously by segmentation over long or short sequences, while the targets may disappear during tracking and reappear later in the video.
VOTS adopts a general problem formulation that covers single/multiple-target and short/long-term tracking as special cases. The tracker is initialized in the first frame by segmentation masks for all tracked targets. In each subsequent frame, the tracker has to report all segmentation masks (one for each target). The following figure summarizes the tracking task.
Follow https://votchallenge.net/howto/overview.html for how to create your submission. Do not forget to pack the results with the
vot pack command.
Make sure that in the tracker identifier in the manifest.yml (by default is inside the output zip file) match with the tracker short name you register through our Google Form.
Then submit your zip file in Participate tab.
Does the number of targets change during tracking?
All targets in the sequence are specified in the first frame. During tracking, some targets may disappear and possibly reappear later. The number of targets is different from sequence to sequence.
Can I participate with a single-target tracker?
Sure, with a slight adjustment. You will write a wrapper that creates several independent tracker instances, each tracking one of the targets. To the toolkit, your tracker will be a multi-target tracker, while internally, you’re running independent trackers. See the example here.
Can I participate with a bounding box tracker?
Sure, with a slight extension. In previous VOT challenges we showed that box trackers achieve very good performance on segmentation tasks by running a general segmentation on top of a bounding box. So you can simply run AlphaRef (or a similar box refinement module like SAM) on the top of your estimated bounding box to create the per-target segmentation mask. Running a vanilla bounding box tracker is possible, but its accuracy will be low (robustness might still be high).
Which datasets can I use for training?
Validation and test splits of popular tracking datasets are NOT allowed for training the model. These include: OTB, VOT, ALOV, UAV123, NUSPRO, TempleColor, AVisT, LaSOT-val, GOT10k-val, GOT10k-test, TrackingNet-val/test, TOTB. Other than above, training splits of any dataset is allowed (including LaSOT-train, TrackingNet-train, YouTubeVOS, COCO, etc.). For including the transparent objects, it is allowed to use the Trans2k dataset.
Which performance measures are you using?
New performance measures are developed for the VOTS challenges, here is a draft.
When will my results be publicly available?
The results for a registered tracker are immediately revealed to the participant via an email. But these results will not be publicly disclosed until after the VOTS2023 workshop. At that point, a public VOTS2023 challenge leaderboard will appear on the VOTS webpage.
If I submit several time to the evaluation server, which submission will be used for the final score?
The final submission will be used for the final score. Please make sure that the tracker description matches the code that produced the final submission.
Questions regarding the VOTS2023 challenge should be directed to the VOTS2023 committee. If you have general technical questions regarding the VOT toolkit, consult the FAQ page and the VOT support forum first.
Start: March 1, 2023, midnight
Description: Result submission
June 18, 2023, 4:59 p.m.
You must be logged in to participate in competitions.Sign In