问题
Let say I want to train a HOG descriptor + Linear SVM for a car detection. Is it necessary for me to make, let say three classifiers, that are back-view, front-view and side-view of the car or I can just train a single classier for all viewpoints of the car?
回答1:
It's not necessary but recommended. You can make a single classifier which handles multiple cases but it won't perform very well overall. The issue here isn't so much the variability of descriptor responses between the different views, but the difference in aspect ratios between rear/front-facing and side-facing detectors. The sliding window you use to extract your HOG will either capture too much negative (sideview-sized over rear/front) or not enough positive data (rear/front-sized over sideview).
Bottom line: it depends on your accuracy/processing rate requirements. From my experience front/rear are usually similar enough, but if high accuracy is desired, you'll need a separate detector for each. You will definitely need a separate side detector and you might even need dedicated 'quarter view' detectors - front right, front left, rear right, rear left.
P.S. I'm omitting the fact that you will also need to handle multiple scales.
来源：https://stackoverflow.com/questions/37671278/computer-vision-is-it-necessary-to-have-multi-classifiers-with-certain-viewpoi