DAVIS (Densely Annotated VIdeo Segmentation) is an academic dataset and evaluation benchmark for video object segmentation. The site states that its goal is to provide in-depth analysis of state-of-the-art methods in video object segmentation, and it offers entry points for DAVIS 2016, DAVIS 2017, and past Challenges. Note that the standalone DAVIS project is now in maintenance mode: no new DAVIS Challenges will be held, and the benchmark leaderboard on the site is no longer updated. Existing results have been integrated into Papers with Code.
DAVIS 2016 annotates a single instance per video, while DAVIS 2017 includes multi-instance annotations and distinguishes between Semi-supervised and Unsupervised settings. Here, semi-supervised/unsupervised refers to the degree of human interaction at test time: semi-supervised provides the target mask in the first frame, while unsupervised provides no human input. The site also offers State-of-the-Art result displays, in-browser result visualization, image and annotation downloads, precomputed results, and code for reproducing evaluations. The evaluation server is still running on Codalab, and results from new papers can be submitted to the corresponding task pages on Papers with Code.
The main text does not mention fees, subscriptions, commercial licensing, or payment methods, so DAVIS is better understood as a free and open research benchmark resource rather than a commercial SaaS product. It is not a typical developer tool platform: there is no mention of APIs, SDKs, CLIs, cloud services, or self-hosted deployment options. Supported programming languages and deep learning frameworks are also not specified in the main text. However, the availability of reproducible evaluation code and public dataset downloads still makes it useful for research-oriented development workflows.
Its strengths are a clearly defined dataset, a well-established citation system, complete CVPR/arXiv papers and BibTex entries from previous years, plus community-extended annotations such as referring expressions, eye-gaze, and shadow annotations. It is well suited to research in video object segmentation, visual attention, and multimodal segmentation. Its drawbacks are limited update activity since the project entered maintenance mode, the discontinued official leaderboard updates, and a lack of support services, licensing details, and engineering-oriented documentation. DAVIS is best suited for computer vision researchers, paper authors, and algorithm teams that need standardized VOS evaluation.
The main text does not provide information about access from mainland China, mirrors, or payments, so its access status can only be marked as unknown. If access to Codalab, Papers with Code, or dataset downloads is unstable, researchers may consider using similar video object segmentation datasets such as YouTube-VOS, MOSE, and LVOS, or directly checking the corresponding task pages on Papers with Code for updated results and alternative benchmarks.
โ This review is compiled from public sources and does not constitute a purchase recommendation. Verify all facts on the vendor's official site. Verify on davischallenge.org official site.
davischallenge.org is an Switzerland Dev Tools provider. TG4G tracks its product information, an overall rating of 8.0/10, and a China-accessibility score of China direct-connect friendly. Click "Visit Official Site" to reach davischallenge.org directly.