Hello,
I am planning to work on the Ways to improve your facial recognition model : mention on the face recognition assignment
Can i use yad2k for face detection or do I have to use Yolo3? Please advise
Looks like you should be able to:
In my opinion the question of which version of YOLO to use, or for that matter whether to use any version of YOLO, has to be driven by the project requirements, which you haven’t shared. To the best of my knowledge, the principle improvement in v3 is explicit training at multiple scales, which isn’t done in YOLO 9000 aka v2. Is that important in your application? It also may matter what you’re going to deploy on. Some releases of YOLO took more consideration of constrained (read mobile) operating environments. My implementation of v2 has over 50 million trainable parameters. The model is big. Additionally, what initially got people excited about YOLO was prediction-time speed. YOLO has never been the most accurate, but it is reasonably accurate and very fast compared to other algorithms. Some situations demand this, whereas in others you can sacrifice a few (or many) frame per second throughput for higher precision. YOLO is also designed specifically to address the need to detect multiple objects per input. You can do more, better, faster with a simpler algorithm if you only have a single object to deal with. Another consideration is training data itself. My recent experiments on training YOLO from scratch suggest that it takes a significant amount of congenial data to train a model. Do you have that available? You should also plan ahead for substantial computational resources for training. Not trying to dampen your enthusiasm here, or discourage you from diving deeper into YOLO. But solution architecture is a set of technical means. Unless this is a purely academic experiment, I encourage you to ensure the ones you select are motivated by and directly aligned with business ends.
ps: Note that the zelener archive referenced above is 5 years old. It is not the only YOLO v2 python/tf/keras implementation out there now, though it may have been when this course was initially developed. In my opinion after comparing a few, it is not the easiest to read or modify. Do yourself a favor and at least take a look at some of the others for ideas.