Question about evaluation on custom dataset

[man-droid23]

Question about evaluation on custom dataset

[kevang33]

I have something similar, the 3d results are much lower even after several epochs of training. I put my data into KITTI format, then trained a pointpillars model on it, but the boxes that are outputted from the model don't seem right. It seems that the coordinates of the predicted boxes may be flipped - the y coordinates of the predicted boxes are more similar to the x coordinates of the ground truth boxes

Here are what the predicted boxes look like after running demo.py on my trained model:

And here is what some of the ground truth boxes look like:

@man-droid23 have you solved this issue?