DAVIS (Densely Annotated VIdeo Segmentation) is one of the most important datasets in the video object segmentation (VOS) task, and its purpose is to conduct in-depth analysis of the state-of-the-art methods in the field of video object segmentation.
Purpose DAVIS is aimed at the video object segmentation task, providing densely annotated video data and accompanying evaluation methods for training and evaluating object segmentation models in videos.
Settings and Versions DAVIS evaluation is divided into two categories: semi-supervised and unsupervised, where "semi-supervised" and "unsupervised" here refer to the degree of human interaction in the testing phase, not the training phase. Semi-supervised (more precisely, human-guided) provides the segmentation mask of the object of interest in the first frame; unsupervised (more precisely, not guided by humans) does not provide any human input. In terms of versions: in DAVIS 2016, each video sequence annotates a single instance; in the DAVIS 2017 semi-supervised setting, each video sequence annotates multiple instances; in the DAVIS 2017 unsupervised setting, each video sequence likewise annotates multiple instances.
Main Capabilities DAVIS provides densely annotated data, benchmark evaluation, and displays of results from current state-of-the-art methods. The website can display evaluations of current state-of-the-art segmentation techniques using the three metrics proposed in its work; it supports visualizing in the browser the segmentation results of all current state-of-the-art techniques on all DAVIS 2016 images; and it allows downloading DAVIS images and annotations, precomputed results of all techniques, as well as code for reproducing the evaluation.
Related Papers The 2019 DAVIS Video Object Segmentation Challenge: Unsupervised Multi-Object Segmentation (S. Caelles, J. Pont-Tuset, F. Perazzi, A. Montes, K.-K. Maninis, L. Van Gool, arXiv:1905.00737, 2019). The 2018 DAVIS Video Object Segmentation Challenge (S. Caelles, A. Montes, K.-K. Maninis, Y. Chen, L. Van Gool, F. Perazzi, J. Pont-Tuset, arXiv:1803.00557, 2018). The 2017 DAVIS Video Object Segmentation Challenge (J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, L. Van Gool, arXiv:1704.00675, 2017). Benchmark datasets and evaluation methods for video object segmentation (F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, A. Sorkine-Hornung, CVPR 2016).
Community Contributions The community has extended the DAVIS dataset: — Referring expression annotations for DAVIS 2016 and DAVIS 2017: includes referring expressions given by two annotators for DAVIS 2016 and DAVIS 2017 when the first frame is given; in addition, it also includes referring expressions given by two annotators for DAVIS 2017 when the entire video is given. The related paper is "Video Object Segmentation with Language Referring Expressions" by A. Khoreva, A. Rohrbach, B. Schiele (ACCV 2018). — Human eye-tracking gaze data for DAVIS 2016: collected from twenty participants using professional equipment, covering all frames of the DAVIS 2016 training and validation sets. The related paper is "Learning Unsupervised Video Object Segmentation through Visual Attention" by W. Wang, H. Song, S. Zhao, J. Shen, S. Zhao, S.C.H. Hoi, H. Ling (CVPR 2019). — Shadow annotations for DAVIS 2016: includes object shadow annotations for 29 videos in DAVIS 2016. The related paper is "Temporally Coherent Completion of Dynamic Video" by J. Huang, S. B. Kang, N. Ahuja, J. Kopf (ACM 2016).
Operating Status The independent DAVIS project is currently in maintenance mode: no DAVIS challenges are held anymore, and the benchmark leaderboard on this page is no longer updated. Existing results have been integrated into paperswithcode, where users can enter the results of the latest papers on the corresponding task pages (semi-supervised, interactive, unsupervised). The evaluation server will continue to run on Codalab. The project team thanks the community for participating in the challenge and encourages everyone to continue using the dataset for video object segmentation or other tasks.
Limitations Apart from the content already provided above, facts about the dataset's specific scale, image resolution, license agreement, charging method and price, platform support, etc., are not yet available in the materials.
