search

gsig / actor-observer

ActorObserverNet code in PyTorch from "Actor and Observer: Joint Modeling of First and Third-Person Videos", CVPR 2018
GNU General Public License v3.0
76 stars 9 forks source link

Why one frame belongs to different actions? #9

Open yuhuangyue opened 4 years ago

yuhuangyue commented 4 years ago

Hi~ In the Charades and CharadesEgo dataset, one video always contains several actions. In the code, you divide the video into several clips according to the start and end time, but I have observed that one frame may belongs to multiple action tags. In this case, Can the loss function be trained normally?

gsig commented 4 years ago

The simplest way of training, using a 1-of-N loss, just considers clips independently like you described. More recent repos (like GitHub.com/gsig/PyVideoResearch) implement sigmoid training too, which is what the best methods now use on the Charades dataset.

Hope that helps!

On Thu, Dec 19, 2019, 2:40 PM yuhuangyue notifications@github.com wrote:

Hi~ In the Charades and CharadesEgo dataset, one video always contains several actions. In the code, you divide the video into several clips according to the start and end time, but I have observed that one frame may belongs to multiple action tags. In this case, Can the loss function be trained normally?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/gsig/actor-observer/issues/9?email_source=notifications&email_token=AA3OMNMZYQFWLY6F7OZHQ63QZOBU3A5CNFSM4J5H37F2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4IBVEYYQ, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA3OMNP6PGHYH2356OWQ2VDQZOBU3ANCNFSM4J5H37FQ .