TAO Open-World

The "Tracking Any Object in Open-World CVPR 2023 Challenge" consists of two sub-challenges: (i) long-tail challenge and (ii) open-world challenge. These challenges are based on the BURST Benchmark [1], which in turn is an extension of the Tracking Any Object (TAO) dataset [2] that involves pixel-precise segmentation masks for all objects. Submissions for both challenges will be evaluated on three sets of object classes: 1) a 78 "common" class set which roughly corresponds to the 80 standard COCO classes [3]. 2) a 404 "uncommon" class for which there are often very few samples in the dataset. 3) The union of the above two, i.e. the 482 class "all" set. Note that our 482 classes are a subset of the much larger class set for the LVIS dataset for image-level instance segmentation [6]. For open-world tracking benchmark, models can be trained using only labels for the 78 common class set. Training on COCO and/or ImageNet is permitted, as is any kind of self-supervised pre-training on other datasets. For this challenge, we will require code to be submitted for internal revision to ensure models are not trained using labels beyond the permitted datasets. Submissions will be evaluated using OWTA, which is basically a recall-based variant of HOTA that does not penalize false positives. The OWTA metric is computed for each of the three above-mentioned class sets and is denoted by OWTA_com, OWTA_unc and OWTA_all, respectively. For more details about the metrics, please refer to the HOTA metrics paper [4] and Sec. 6 of the BURST benchmark paper [2] This challenge is a part of CVPR 2023 workshop: "Tracking and Its Many Guises: Tracking Any Object in Open-World".

Training Set

Test Set

Sample	Name	FPS	Resolution	Length	Tracks	Boxes	Density	Description	Source	Ref.
	BURST_test	30	1280x720	52194 (29:00)	7963	167132	3.2	The test set of BURST.	link	[1]
	Total			52194 frm. (1740 s.)	7963	167132	3.2

Download