3d inference result problem

[TowerPro]

i have rewrite the code to CPP so the YOLO3D can runs on nVidia Jetson NX Xavier, but in the car class, sometimes the inference result is not good, I think this problem is caused by least squares linear regression(numpy api: np.linalg.lstsq in Math.py), is there any better algorithms to improve this problem? or if is necessary to change the cam_to_file(camera calib file) to my own camera?

the result is same as #13

[LordonCN]

Did u use your own datasets to train it? Actually it works well in my device.

[TowerPro]

Did u use your own datasets to train it? Actually it works well in my device.

I have trained on Kitti Dataset with resnet18, I will try VGG16 instead

[upper127]

I'm using resnet to reason about the 3d bounding box doesn't seem to work very well, the bounding box is all deformed