Disentangling Evidence Modality in Video Classification
More human-understandable interpretation for 3d CNN video classifiers
A universally applicable injection to 3d CNN video classifiers that allows for explainability tools such as grad-cam to distinguish between static (single frame) evidence and moving evidence (motion) that contributed to model decision. Easy fine-tuning and no effect on accuracy in experiment.
Applied to Clevelend Clinic’s cardiac AI ejection fraction classifation models.
