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

Most action recognition models today are highly parameterized, and evaluatedon datasets with appearance-wise distinct classes. It has also been shown that2D Convolutional Neural Networks (CNNs) tend to be biased toward texture ratherthan shape in still image recognition tasks, in contrast to humans. Takentogether, this raises suspicion that large video models partly learn spuriousspatial texture correlations rather than to track relevant shapes over time toinfer generalizable semantics from their movement. A natural way to avoidparameter explosion when learning visual patterns over time is to make use ofrecurrence. Biological vision consists of abundant recurrent circuitry, and issuperior to computer vision in terms of domain shift generalization. In thisarticle, we empirically study whether the choice of low-level temporal modelinghas consequences for texture bias and cross-domain robustness. In order toenable a light-weight and systematic assessment of the ability to capturetemporal structure, not revealed from single frames, we provide the TemporalShape (TS) dataset, as well as modified domains of Diving48 allowing for theinvestigation of spatial texture bias in video models. The combined results ofour experiments indicate that sound physical inductive bias such as recurrencein temporal modeling may be advantageous when robustness to domain shift isimportant for the task.

Recur, Attend or Convolve? On Whether Temporal Modeling Matters for Cross-Domain Robustness in Action Recognition