Any suggest about improving different viewpoint detection performance with additional data？ about handtracking HOT 2 CLOSED

Great question!

I honestly do not know enough to give a confident answer as I have not run many similar experiments. However, my intuition is that if your dataset contains vastly different images (e.g the hands in some samples differ greatly from the hands in others) for same class, the network may have a hard time learning the right features that capture all of the data. This might mean the network may take longer to train and the final loss values may not be very low (or indicate a failure to converge).
From this point of view, it may be useful to have similar image size/resolution/quality. The more the data the better, so having comparable or more non-egocentric data should yield good results.

I would say you run some experiments and please share the results if you can!
-V.

from handtracking.

OK, Thanks, I will try!

from handtracking.