I found that in training stage, the frame number of every videos (frame groups) are random. I wonder if this variable length training strategy is better than the fixed length training strategy.
The pseudo code of fore-mentioned fixed length training strategy is shown below.
for i in range(epoch):
for scene in scene_list: # In NuScenes dataset, frame length of every scenes are approximately equal (around 40).
for frame in frame_list:
....
train(data[scene][frame])
....
Thank you for this great work!
I found that in training stage, the frame number of every videos (frame groups) are random. I wonder if this variable length training strategy is better than the fixed length training strategy.
The pseudo code of fore-mentioned fixed length training strategy is shown below.