Performance

[emilharutyunyaNN]

I am working on motion recognition, and happened to use DDNet as a baseline. I keep getting around 0.72 performance with the pytorch code, on JHMDB dataset. The difference in my code is that I use sapiens pose estimator to estimate the keypoints of the frames then later do action recogntion, because of memory shortage I divide each video into 2 16 frame snippets and then pass the predicted keypoints of those to the DDNET processing and training pipeline. Could this be the part causing the lower performance? I would appreciate if you shared, what bug got fixed.