custom training of MOTMPNet

[philip-fu]

Hi,

I am trying to train the model with pre-computed detections and embeddings because my detector is not a faster-rcnn like Tracktor and my CNN is not based on torchreid. Just checking if this will work,

1. use the gt boxes and compute embedding features and store in some folder with some format.
2. modify MOTGraphDataset class so it doesn't compute again

Let me know your thoughts. Thank you!