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 recent advances in video action recognition achieving strongperformance on existing benchmarks, these models often lack robustness whenfaced with natural distribution shifts between training and test data. Wepropose two novel evaluation methods to assess model resilience to suchdistribution disparity. One method uses two different datasets collected fromdifferent sources and uses one for training and validation, and the other fortesting. More precisely, we created dataset splits of HMDB-51 or UCF-101 fortraining, and Kinetics-400 for testing, using the subset of the classes thatare overlapping in both train and test datasets. The other proposed methodextracts the feature mean of each class from the target evaluation dataset'straining data (i.e. class prototype) and estimates test video prediction as acosine similarity score between each sample to the class prototypes of eachtarget class. This procedure does not alter model weights using the targetdataset and it does not require aligning overlapping classes of two differentdatasets, thus is a very efficient method to test the model robustness todistribution shifts without prior knowledge of the target distribution. Weaddress the robustness problem by adversarial augmentation training -generating augmented views of videos that are "hard" for the classificationmodel by applying gradient ascent on the augmentation parameters - as well as"curriculum" scheduling the strength of the video augmentations. Weexperimentally demonstrate the superior performance of the proposed adversarialaugmentation approach over baselines across three state-of-the-art actionrecognition models - TSM, Video Swin Transformer, and Uniformer. The presentedwork provides critical insight into model robustness to distribution shifts andpresents effective techniques to enhance video action recognition performancein a real-world deployment.

Adversarial Augmentation Training Makes Action Recognition Models More Robust to Realistic Video Distribution Shifts