conclusion.tex

\section{Conclusion}
Deep learning has greatly improved the performance of video objects segmentaion tasks. As the previous work shown, firstly we need to design more and more powerful network
to have rich feature representation. In a word , we need to combine the low-level feature and high-level feature. Then temporal information is another important
cue when tackling  video dataset. We will benefit performance improvements by combining temporal information. There are still huge space for us to explore how 
to utilize the temporal information well.