fmassa/object-detection.torch

Is there code for viewpoint estimation?

yw411 opened this issue · 4 comments

yw411 commented

I read the paper 'CRAFTING A MULTI-TASK CNN FOR VIEWPOINT ESTIMATION' and find this pository which is used for object detection,and I saw the result of examples. So I want to know whether the text on the top left corner is the result for viewpoint estimation,including class and azimuthal angle?

fmassa commented

Yes, there is a preliminary release of the code that can be found in http://imagine.enpc.fr/~suzano-f/bmvc2016-pose/ and is based on the code in this repo.

ardiya commented

Hi @fmassa, I'm also here for CRAFTING A MULTI-TASK CNN FOR VIEWPOINT ESTIMATION.

I tried to load Pascal3D dataset using your preliminary release, however the format seems to be different; It reads .xml annotation rather than .mat annotation on P3D dataset.

Could you share us the DataSetPascal.lua used in your paper?

fmassa commented

Hi @ardiya,
Sorry for the (long) delay in replying.

I have used the code I posted in the zip file for training the models. If the original annotations for Pascal3D are in .mat format, I probably converted them to .xml beforehand.
I'll try to have access to the computer where I have stored that data and see if the information is still there, and to see if I still have the conversion script somewhere.

fmassa commented

@ardiya I found the Matlab code I used to convert the .mat files to .xml. It requires the toolbox from Pascal VOC to be on your path.
The code is not very clean but should be easy to use.
You can find it in here https://gist.github.com/fmassa/f9f39bb07cca98d5a1b57615b4c0ea60