Action recognition. The Pascal VOC 2012 action recog- nition task consists of 4588 training images and 4569 test images featuring people performing actions among ten cate- gories such as jumping, phoning, playing instrument or reading. This fine-grained task differs from the object classification task because it entails recognizing fine differences in human poses (e.g. running v.s. walking) and subtle interactions with objects (phoning or taking photo). Training samples with multiple simul- taneous actions are excluded from our training set.To evaluate how our transfer method performs on this very different target task, we use a network pre-trained on 1512 ImageNet object classes and apply our transfer methodology to the Pascal VOC action classification task.