search

zhoubolei / TRN-pytorch

Temporal Relation Networks
http://relation.csail.mit.edu/
Other
787 stars 190 forks source link

putIterationsPerSec before added to Buffer #87

Open martinpflaum opened 4 years ago

martinpflaum commented 4 years ago

Hi i am really a beginner so i am not sure if i got this right, but inn line 136 of fps_dem_trn.py putIterationsPerSec is called which will draw the class label on this image. But the same image is also used to predict the labels later on because it is put into the buffer. Why not switch line 140 and 136 and then instead of calling crop_img = frame[0:720, 0:720] do crop_img = frame.copy(). Also line 144 will be needed to be changed to cv2.imshow("Frame", frame). And thanks for the great repo its really cool to see the model working in realtime.

Regards Martin

martinpflaum commented 4 years ago

Its also a bit strange that there is no regulation in which frequence the frames are put into the buffer. so accuracy changes depending on how fast or slow the pc is. Does the framerate on which the model was trained differ a lot from 12 fps (Jester Dataset)? regards martin