Fu-Hsiang Chan comments

Results 15 comments of


                                            Fu-Hsiang Chan

Difference between your implementation and the actual paper's?

Hi @mittalrajat , Yes, you are right.

Question about loss

Hi @e2sh83 , Sorry, I didn't log the training loss. But, I had trained the Deep3DBox only 20 epochs and evaluated the AP of orientation. The following table is the...

Question about loss

Hi @hectorgui, I split the KITTI into train/val set used in 3DOP/Mono3D/MV3D. And my result is testing on the validation set.

Question for the box2d part!

Yes, The KITTI train/val split used in 3DOP/Mono3D/MV3D.

How to get the 3d proposal from 2d?

Hi @zhaobaozi I only implemented the CNN part of the paper and you only can get the 3D dimension and orientation of 3d proposals. In this project, I didn't implement...

Can't understand the angle trained in network

Hi @sdukaka, It just orientates +pi/2, and it will subtract pi/2 in the final output.

Can't understand the angle trained in network

Yes, the outputs of network are 3D dimensions and observation angle of the object.

3D Box orientation Regression problem

Hi @weisili2016 , The CNN will predict the local object orientation, not the global object orientation. So in testing phase, θlocal will add with θray. https://github.com/smallcorgi/3D-Deepbox/blob/c409d46648afc241930972bae5f1a2c700816862/main.py#L252

3D Box orientation Regression problem

@weisili2016 Yes.