Hi, I didn't notice this reply for a long time, but felt bad not to reply to something so long :) 1. Yes, good datasets exist. They grow over time and they are

Hi, I didn't notice this reply for a long time, but felt bad not to reply to something so long :) 1. Yes, good datasets exist. They grow over time and they are open. It takes a lot of time. ILSVRC is good example. There are many more: Pascal VOC, COCO probably best known among them. 2. The optima vary. Training set has huge impact and is super super important. Has to be large and varied otherwise you've lost even before you choose a model 3. They can be very fast to run. With a good GPU, the most recent convnets can run in few milliseconds per image 4. Everything is open and BSD/MIT, no issues at all. Models are often distributed, for example the ILSVRC 2012 model is included with Caffe framework. 5. The trained models are TRIVIAL for layman to use. But they are not trivial to train. 6. Yes there is a lot of specialized knowledge needed to tweak model params. The summary is these models are now trivial to use - you literally give it an image and it gives you very good predictions for what's inside in a few milliseconds (or few seconds on CPU, not GPU). They are not trivial to train and understand, though. That needs time, practice and mathematical understanding. Caffe is the best framework to look at. Pretrained models are available for 1000 ILSVRC classes for classification and for 200 ILSVRC classes in detection, but not for many other tasks (e.g. scene classification etc.)