He soon realized that the visual information is on its own and is limited, so he needed long-term information, which he got from what the team calls an action history. “We explore three different ways to extract the action history. One, it's from a visual language model, prompting it to obtain what happened in the past. And another way is we reuse the features that we had, the visual features that we had for the anticipation, to perform a recognition, and we start to create an action history. So we use the past information, as in a real-time setting, to build an action history and to encode the actions and fuse 26 DAILY WACV Sunday Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=