|
|
|
|
|
|
by zjj
6097 days ago
|
|
|
Starting with some basic knowledge of machine learning (clustering, NN, bayesian inference, etc.) and some basic computer vision / processing (edge detection, color, basic shapes), how much theory is needed for achieving that objective? (recognizing vehicles in photos, and more interesting objectives: extracting 3d structure from a single 2d image). |
|
|
For the most part, it doesn't matter what classifier you use: k nearest neighbor,support vector machine, random forest, neural nets. They'll all give about the same performance. You should have a general idea what they do, but I don't think it's worth the effort to become a "neural net expert". You should know enough pattern recognition so you don't fool yourself (by over-training, for example), and have an idea for how to choose the right features.
Where should you put your effort? Into finding useful features for the object you want to classify. And the more image processing you know the more useful features you'll be able to try. How much do you need to know? Depends on the problem. If you're finding cars in the desert then not so much. Your feature set might be "has long straight lines and is not sand colored". If you're trying to tell American made cars from Japanese then it's harder (unless they are moving, in which case it can't be American).