Hi,
I have been working on online action recognition (you estimate actions from stream of a video). I have read some papers about this issue. There are two approaches that often utilized.
1. Using sliding window to estimate an action from sequence of frame.
2. Estimating action frame by frame
For the first approach, mostly missclassification occurs during onset and offset phase (when a person starts or stop performing an action).
For the second, we should predict an action based on glimpse of information, since decision should be made as soon as possible.
I want to ask about what do you think about those approaches?
Any suggestions will be very helpful
Nowadays everyone is using a sequence of frames because from a single frame we can not recognize the action accurately. In a single frame, we can predict only the object inside the image but the action itself is a sequence of frames showing the movement of objects or object parts. I would recommend my friend research project of action recognition using bi-directional lstm... https://github.com/Aminullah6264/BidirectionalLSTM
you can study this paper and code hope it will be helpful for you
Nowadays everyone is using a sequence of frames because from a single frame we can not recognize the action accurately. In a single frame, we can predict only the object inside the image but the action itself is a sequence of frames showing the movement of objects or object parts. I would recommend my friend research project of action recognition using bi-directional lstm... https://github.com/Aminullah6264/BidirectionalLSTM
you can study this paper and code hope it will be helpful for you
Thank you for your replies.
Salman Khan : That is interesting paper. And I think in that paper the authors performed trimming of sequence of images from an action. That can be applied in offline scenario (I suppose). In online scenario, we have to tackle another challenge. That is partial information, in which the current sequence that we have sometimes represent transition phase (onset or offset phase) from one action to another action. Recognizing this phase is quite challenging as the information is ambiguous (assume we use sliding window).
Yasser Jaamour : Do you mean to estimate depth information from single image?
Prasetia Utama : precisely yes, it is a new research topic in image and video processing.
- Badges
- Science topic
- Similar topics
- Papers