The issues of adapting the other models to the omni3d datasets

[Ivan-Tang-3D]

Hi, thanks for your great work. When adapting some mono 3D detectors to the omni3d datatset, I encounter the circumstance that the 3D AP remains below 1 after about 20000 iterations. Therefore, I wnoder whether it is possible to release the code of adapting the other models, like SMOKE, to the dataset, just as the paper shows. Meanwhile, I wonder if it is possible to give me some clues about what parts might lead to such results specifically, like the angle loss, the training post process(virtual depth), data augmentation. Thanks very much.