search

scaelles / OSVOS-TensorFlow

One-Shot Video Object Segmentation
http://vision.ee.ethz.ch/~cvlsegmentation/osvos/
GNU General Public License v3.0
438 stars 132 forks source link

Contour Branch #28

Closed msanzn closed 4 years ago

msanzn commented 4 years ago

I have the doubt if this implementation includes the contour branch as running the implementation shows the object detected with very strange borders. From the published paper I understand that the contour branch is only used for the parent training and not for the training done with one mask.

in the image !{00024][logo] it can be seen that the border does not match with the back of the vehicle, so by supposing that in imagenet the vehicles data are not in that position.

I would like to know if there is a way a way to improve the contour matching and the detection on object outside imagenet as posters, paintings or road signs.

00024

scaelles commented 4 years ago

Hello, As mentioned in a previous issue, the contour branch is only available in the Caffe version. The countour branch is trained separately and only joined with the appearance branch at test time. The network is not trained for any specific category and it should be able to segmentat any object category given the mask for that object in the first frame of a video.

msanzn commented 4 years ago

I am triying to use the Fast Bilateral Solver to see the improvement in the my testing frames, in the opencv website I had only found the implementation and example in python. Also it would be great if could explain if you introduced the mask image or if it is necessary some prior conversion.