Our entry stood fourth and we presented our work at the workshop poster, we had utilized a 42 model ensemble (Resnets, Resnexts ) Our top 3 best accuracy on a single model was 89.6 % while with using pretrained feature of penultimate layer of ResNext and then applying xgboost with 0.01 learning rate and 200 rounds , the top 3 accuracy is 86.7 %

• densenet-pytorch
• Wide Residual Networks (BMVC 2016) http://arxiv.org/abs/1605.07146 by Sergey Zagoruyko and Nikos Komodakis.
• BC leanring (CVPR 2018) https://arxiv.org/abs/1711.10284 by Yuji Tokozume, Yoshitaka Ushiku, Tatsuya Harada

Rishabh Gupta, Atsuhiro Noguchi and Kuniaki Saito (The University of Tokyo)