You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[Before trying anything, train 1 model with current state to be able to compare]
[After this comes issue #1]
So far, optical flow is calculated thanks to OpenCV's function cv2.calcOpticalFlowFarneback. In the original paper, DeepFlow is used. This difference might be the source of the poor quality of the Temporal Loss.
Clone DeepFlow github repo
Add it to the requirements
Test DeepFlow on some images and compare it with the current solution
If the results are coherent and better than the current, switch it.
/!\ For better training performances, It might be necessary to compute all dataset's optical flow once and store it (changing the current dataset organisation). This will allow us not to compute optical flow on each example of the training set.
The text was updated successfully, but these errors were encountered:
[Before trying anything, train 1 model with current state to be able to compare]
[After this comes issue #1]
So far, optical flow is calculated thanks to OpenCV's function cv2.calcOpticalFlowFarneback. In the original paper, DeepFlow is used. This difference might be the source of the poor quality of the Temporal Loss.
/!\ For better training performances, It might be necessary to compute all dataset's optical flow once and store it (changing the current dataset organisation). This will allow us not to compute optical flow on each example of the training set.
The text was updated successfully, but these errors were encountered: