I play some object detections for some years, ex: HOG+SVM or darknet yolo algorithm. Now I focus on the rotation invariant and I find there are got very few attention.
I think rotation invariant object detection is important, for example, the bacteria in the microscope may be in any angle, and we have to train the program that, even the bacteria shows in any angle, there are all the same bacteria.
What I expect is something like a sliding-window based detection with the rotation information. But it seems not really such a thing.
there are some answers to do the "rotation invariant object detection" I find in google.
find the target's contour, shape, or moment. Rotate it to fit the traditional rectangle based object detection and calculate the angle. This method does not really answer the question because in general, we even don't know which contour is our target and there are too many contours to be tried.( However, This method seems popular in IC design )
people say that in a neural network, we can just train the sample of our target objects with all the 360 degrees pictures and we can get the model which can detect our target in any degree. However, as my experience in training the hand poses it might result in that: we may need much more additional layers to learn it, or the model just confused and didn't learn that object.
some people use SURF or SIFT to match the feature points so we can match a target's feature points to our template. But this is more like feature points alignment, not the object detection.
some features seem to be "rotation invariant" such as the one in "Rotation Invariant Object Recognition from one training sample" (Yokono and Poggio,2004). But in this paper, it seems that they still need feature points matching to get the correct angle.
In yolo network we can train the (x,y,w,h) in each pixel so we can " you only look once ". We can also do something similar to train the angle to it and get a rotated bounding box like in "Learning a Rotation Invariant Detector with Rotatable Bounding Box (Liu et al.2017). However, the method is really few referenced and lack of verification.
What's the truly state-of-art and well-verified method of these "rotation invariant object detection" method? I'm surprised that the rotation issue is out of spot even the object detection is one of the most popular topics in computer vision.