Temporal Action Detection

Temporal action detection studies how to find and classify action intervals in long, untrimmed videos, supporting tasks such as sports event spotting, surgical phase monitoring, and online alerts. Common approaches use temporal convolutional networks, transformers, or state-space models to combine local motion cues with longer-range context, then predict boundaries and rank candidate events. Research also addresses weak or point-level labels, streaming and efficient inference, and zero-shot or open-vocabulary detection using language and vision-language models; THUMOS14 and ActivityNet-1.3 are common benchmarks. A central challenge is balancing accurate, coherent boundaries and generalization to unseen settings against the cost of dense annotation and computation, while ensuring evaluation reflects event-level performance rather than frame scores alone.