process data
label cnt = 5005
kernels related to bbox:
https://www.kaggle.com/suicaokhoailang/generating-whale-bounding-boxes/comments
https://www.kaggle.com/martinpiotte/bounding-box-model/output
good public ensemble kernels:
https://www.kaggle.com/moradnejad/ensemble-lb-0-868
https://www.kaggle.com/vanshjatana/let-s-blend-0-867
learn how to build and train simaese network,
it is good but takes too much time for training.