English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Despite impressive performance for high-level downstream tasks,self-supervised pre-training methods have not yet fully delivered on densegeometric vision tasks such as stereo matching or optical flow. The applicationof self-supervised concepts, such as instance discrimination or masked imagemodeling, to geometric tasks is an active area of research. In this work, webuild on the recent cross-view completion framework, a variation of maskedimage modeling that leverages a second view from the same scene which makes itwell suited for binocular downstream tasks. The applicability of this concepthas so far been limited in at least two ways: (a) by the difficulty ofcollecting real-world image pairs -- in practice only synthetic data have beenused -- and (b) by the lack of generalization of vanilla transformers to densedownstream tasks for which relative position is more meaningful than absoluteposition. We explore three avenues of improvement. First, we introduce a methodto collect suitable real-world image pairs at large scale. Second, weexperiment with relative positional embeddings and show that they enable visiontransformers to perform substantially better. Third, we scale up visiontransformer based cross-completion architectures, which is made possible by theuse of large amounts of data. With these improvements, we show for the firsttime that state-of-the-art results on stereo matching and optical flow can bereached without using any classical task-specific techniques like correlationvolume, iterative estimation, image warping or multi-scale reasoning, thuspaving the way towards universal vision models.

CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow