Question about labeling each frame series.

[somanshreddy]

I am trying to use a different dataset where each frame is labeled as opposed to each frame series,
How did you label each frame series? Was it given in the dataset directly or did you manually label each frame series?

We ran OpenPose on our dataset which was at 25 fps, so do you recommend we choose our n_steps as 25 itself? In this case what if an activity begins in the middle of a frame series, should it be labeled as activity or not?

[seiffertacfr]

For each video the dataset is based on, the person is only ever performing 1 of 6 activities, so the label would always be the same for an entire sequence.
As the model is currently set up, you only want a single label per input, so one label for the entire frame series, you could preprocess your data to ensure each sequence contains only 1 label, or just take the most common label.
You could change this if you want and have the output try and match a given % of each known activity performed during a sequence. In this way if a sequence contains some frames labelled as one activity and other frames labelled as another activity the input label would be changed to reflect the % of frames per total known labels. You would need to change the loss function for this to work.

I would use a larger sequence than 1 second to train on. The only reason I chose 32 for n_steps was because I wanted a window to cover about 2 seconds, as this seemed to be how long it took me to easily determine what activity each sequence contained in the videos, you could definitely test shorter input sequences on your data, but there is no reason it should be the same as your fps.