In paralell to MSRA’s release of their winning architecture of the ILSVRC 2015 researchers from Google have released a substantial improvement of their inception architecture (their winning model of ILSVRC 2014, improved in Feb 2015 by the introduction of batch normalization). The improved architecture is called inception v3 and is on par with MSRA’s winning architecture in terms of top-5 / top-1 multi-crop error on the ImageNet validation set, see Rethinking the Inception Architecture for Computer Vision. An interesting blog-post from researchers from Facebook and Cornell Tech seems to indicate that inception v3 is more efficient than deep residual networks in terms of their accuracy / ms ratio (ms relates to the milliseconds of a full forward – backward pass on a 32 mini-batch whereas accuracy is the top-1 single crop error on the ImageNet validation set), see Training and investigating Residual Nets. The pre-trained model is available e.g. in tensorflow (Tensorflow – Imagenet-InceptionV3) and mxnet (DMLC-mxnet pretrained models).