search

feinanshan / TDNet

Temporally Distributed Networks for Fast Video Semantic Segmentation
http://cs-people.bu.edu/pinghu/TDNet
MIT License
201 stars 45 forks source link

NYUD less training images #18

Closed brdav closed 4 years ago

brdav commented 4 years ago

Hello, thanks a lot for providing the code and datasets! I noticed that the NYUD training set only contains 780 instead of the usual 795 samples. What is the reason for leaving out the other 15? Best David

feinanshan commented 4 years ago

Hi, this is due to the failures that matching the image to video frames. For NYUD, we located the video clips by comparing the difference between the rectified video frames and the annotated images. The 15 training samples may be missed due to the failures of matching. This doesn't significantly affect the results and conclusion. We also suggest directly replicating the 15 images to form 15 still videos, so as to achieve a complete dataset.