As also pointed out in #43 , for both training and testing the calibration matrices used are always the averaged one instead of the correct ones for each single frame.
In different frames this creates a visible offset that might hinder the training process.
The modifications here implemented address this problem: first fetching the correct calibration matrices P, T and R, then passing them down to the functions used (mainly in utils.py)
As also pointed out in #43 , for both training and testing the calibration matrices used are always the averaged one instead of the correct ones for each single frame. In different frames this creates a visible offset that might hinder the training process.
The modifications here implemented address this problem: first fetching the correct calibration matrices P, T and R, then passing them down to the functions used (mainly in utils.py)