I currently have a problem where I need to isolate certain parts of a bicycle from a static image.........of a bicycle. Initially, I had tried using Hough circle transforms to detect the wheels and go from there, however this produced poor results when wheels were skewed into ellipse shapes (which is almost always the case). For clarity, I am using Python 3.6 and OpenCV (obviously).
I am interested to know how others would go about solving this problem? I am considering collecting HOG descriptors from positive images of bicycle wheels and using them to train an SVM classifier. Would this be able to account for the variance in the wheels skew and scale? Using a deformable part based model to recognise parts of the bicycle would be ideal, however this isn't included in OpenCV 3.1 unfortunately.
Because the problem is limited to static images, I don't really need the efficiency that most online resources look for (most resources apply object detection to videos). I just need an accurate way of locating the areas of interest on a static image of a bicycle. The challenging part of this is the fact that a bicycle is an awkward shape and is mostly not opaque. Furthermore, there is a great deal of intra-class variability for a bicycle object.