image-to-3d-bbox
image-to-3d-bbox copied to clipboard
OK, don't waste time on this project
I don't think the author reproduce the result, I try many method just can not got any result. Reasons are:
-
The model can not predict dimensions and angle properly. I trained the model, and the result predicted are all the same value (every object are same dimension and almost same angle which means the network is dead);
-
Even we got the dimensions and angle, we can not get center of 3d box, because this project using label file that contains 3d center as already known.
If anyone wanna regress 3d box, better try another project or another method, this method just not work....
Have you failed to predict the dimension and even the orientation? What were issues? Do you also know another project similar to this?
@jinfagang You can try this project 3D_detection, the owner added center predict method (which still needs optimization).
i was puzzle,this is not correct. it is so easy and different from ssd and rccn