Benchmarking the Robustness of Spatial-Temporal Models Against Corruptions

Robustness
DOI: 10.48550/arxiv.2110.06513 Publication Date: 2021-01-01
ABSTRACT
The state-of-the-art deep neural networks are vulnerable to common corruptions (e.g., input data degradations, distortions, and disturbances caused by weather changes, system error, processing). While much progress has been made in analyzing improving the robustness of models image understanding, video understanding is largely unexplored. In this paper, we establish a corruption benchmark, Mini Kinetics-C SSV2-C, which considers temporal beyond spatial images. We make first attempt conduct an exhaustive study on established CNN-based Transformer-based spatial-temporal models. provides some guidance robust model design training: performs better than robustness; generalization ability implies against corruptions; (especially domain) enhances with computational cost capacity, may contradict current trend efficiency Moreover, find intervention for image-related tasks training noise) not work
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES ()
CITATIONS ()
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....