One-shot action recognition in challenging therapy scenarios

One-shot action recognition aims to recognize new action categories from asingle reference example, typically referred to as the anchor example. Thiswork presents a novel approach for one-shot action recognition in the wild thatcomputes motion representations robust to variable kinematic conditions.One-shot action recognition is then performed by evaluating anchor and targetmotion representations. We also develop a set of complementary steps that boostthe action recognition performance in the most challenging scenarios. Ourapproach is evaluated on the public NTU-120 one-shot action recognitionbenchmark, outperforming previous action recognition models. Besides, weevaluate our framework on a real use-case of therapy with autistic people.These recordings are particularly challenging due to high-level artifacts fromthe patient motion. Our results provide not only quantitative but also onlinequalitative measures, essential for the patient evaluation and monitoringduring the actual therapy.