Microsoft Research Asia (MSRA) is winning both Imagenet Detection and Localization Challenge 2015 as well as MS COCO Detection Challenge 2015

Using very deep convolutional neural nets with relatively low complexity together with region proposal networks (Faster RCNN) MSRA wins both Imagenet Detection and Localization Challenges 2015 and MS COCO Detection Challenge 2015 (ILSVRC 2015MS COCO 2015). By inserting shortcut connections it is becoming feasible to learn extremely deep networks (up to 152 layers) succesfully and improve the error when compared to less deep networks (Exposition of MSRA’s winning model). This approach gives an Imagenet test set top-5 error of 3.57% after ensembling (4.49% for single-model) as compared to 6.66% for last years winning entry “GoogleNet”.

Leave a Reply

Your email address will not be published. Required fields are marked *