"it's starting to become clear that the future power of deep learning is going to come with its own self-contained software package like Caffe or Torch, and not from a dying breed of all-around tool-belts like OpenCV or Matlab. When you share creations made in OpenCV, you end up sharing source files, but with the Deep Learning toolkits, you end up sharing your pre-trained networks. No longer do you have to think about a combination of 20 "little" algorithms for your computer vision pipeline -- you just think about which popular network architecture you want, and then the dataset. If you have the GPUs and ample data, you can do full end-to-end training. And if your dataset is small/medium, you can fine-tune the last few layers. You can even train a linear classifier on top of the final layer, if you're afraid of getting your hands dirty -- just doing that will beat the SIFTs, the HOGs, the GISTs, and all that was celebrated in the past two decades of computer vision."