The calculation method of global feature seems to be different from the SAT paper

[zhouweii234]

The calculation method of global feature gml_feature (line 193 in video_analyst-master/videoanalyst/pipeline/segmenter_impl/sat_pipeline.py) seems to be different from the SAT paper. It did not use state score but only use confidence score to calculate global feature, and seg_init_feature always account for 50% in the global feature.

Is this because I misunderstood the code or this part of the code does not calculate global features? Or is the part of your code that calculates global features really different from the SAT paper?

[lzx1413]

We actually adopt 50% of the init feature for the final segmentation feature. You can refer to the version in this repo.