when choosing current image and other neighbor image to train dff, why set the probability of choosing current frame and other neighbor frame the same?

tonysy / Deep-Feature-Flow-Segmentation

Deep Feature Flow for Video Semantic Segmentation

MIT License

33 stars 4 forks source link

when choosing current image and other neighbor image to train dff, why set the probability of choosing current frame and other neighbor frame the same? #5

Closed Gus-Guo closed 5 years ago

Gus-Guo commented 5 years ago

when choosing current image and other neighbor image to train dff, why set the probability of choosing current frame and other neighbor frame the same? As written in image.py(224)

littletomatodonkey commented 5 years ago

He sets the test method to test the mIOU with different time intervals. But when training the network, it needs to learn many time intervals' impact to the nwtwork to improve the network's generalization ability. I hope this can help you.

tonysy commented 5 years ago

I follow the training strategy of the original paper, @littletomatodonkey has given a good explanation.

Gus-Guo commented 5 years ago

Thanks. I got it.