I COME BACK, hahaha

[Patickk]

sorry ...come again, ..can you give the explain: how to get T ? if we don't get location T(tx, ty, tz), we can't get the rotation_y. Help me ...

[fuenwang]

I just saw it now haha. What do you mean? The orientation our model predict can be directly converted to rotation_y. So we will use it to solve T.

[lucasjinreal]

@fuenwang How to convert predict to rotation_y?

[fuenwang]

@jinfagang https://github.com/fuenwang/3D-BoundingBox/blob/master/Eval.py#L70 This line is the rotation_y the model predict.

[lucasjinreal]

@fuenwang The model predict is actually the local angle of object? Then How do I get the angle what I need to draw a 3d bounding box?

[fuenwang]

The theta in that link is the theta(red) in this figure, so it's already the global rotation of object, not local angle and it is the angle you need to plot 3D box.

[lucasjinreal]

@fuenwang This is global angle, then the camera see object angle is local angle? I am so confused about thoese angles.

In the label file, what alpha stands for? it's the camera see angle? How to convert the label alpha for training label which predicts as the red angle?

[fuenwang]

1. Yes, camera sees local angle.

2. Actually I'm not familiar with alpha either haha, but as far as I know, we can also infer local angle by it. The tensorflow implementation https://github.com/smallcorgi/3D-Deepbox/blob/master/data_processing.py seems using alpha to infer local angle, but I'm not sure how is this working. Maybe you can go through his code if you are interested in it.

[lucasjinreal]

@fuenwang This project I have already tried, it doesn't work at all, the whole network doesn't converge and the ouput result is far away from precise. All the output seems the same and the alpha value can not be used.

Did your code can get precise dimensions and global angle?

[fuenwang]

But his code got much more stars than mine haha.

My code can converge, you can check this closed issue https://github.com/fuenwang/3D-BoundingBox/issues/4 My final angle and dimension error is also in it, I think the error is reasonable.

[lucasjinreal]

@fuenwang So you are actually using rotation_y in label as training labels rather than using alpha and convert it into a global angle?

Then your predict value is actually already the rotation_y, why still using centers to add a arctan of x and z coordinates?