On s3dis, those results are already behind SOTA

[tchaton]

https://arxiv.org/pdf/1904.02113.pdf

[EnyaHermite]

Hi, segments similar to super-pixels cast constraint of local labelling consistency. That's why SOTA or CRF-like networks sometimes outperform ours. However, you might notice that the 6-folds results of SOTA is actually on-par to ours. Compared to the computation cost of SOTA, the efficiency, scalability and generalisation of our network make it potentially more promising.