Zengyi Qin

The `train.py` will start with evaluation ( evaluates at step 0, step eval_iter, step eval_iter * 2 , and so forth). A single image will take

yes, there should be a calib file. I would suggest you to resize (or crop out a preferred region and resize) the image to height 360 * width 1200 before...

Another way that might solve the memory issue: if training is not required, I would suggest to comment out some lines in the `run_training` function in `include/tensorvision/train.py`. Comment the lines...

Hi, thanks for your interest. See https://github.com/Zengyi-Qin/MonoGRNet/issues/32

Hi, thanks for your interest. For each predicted 3D box, we find its nearest GT 3D box. Then compare the two boxes to obtain the errors

The official GT on KITTI dataset are in `data/kittiBox/training/label_2/*.txt`

Thank you for your interest! I would suggest to train as follows: 1. joint_2d_3d = True, others = False, for 80K iterations, Adam. 2. joint_2d_3d = True, others = False,...

In the paper we wrote our original training procedural. But the abovementioned procedural is simpler and can also work

Thanks for your interest! We only have pretrained model on tensorflow. The `setup.py` downloads it.

Thanks for your interest! We did not experiment on other object classes. But [the code](https://github.com/Zengyi-Qin/MonoGRNet/blob/3bf1d6dd0693617cf0eec95ab3d52050f23d48e1/inputs/kitti_input.py#L148) specifies the object classes. You may directly modify the code.