Attempting to create a “shoe detector” that will track and detect a shoe (with a raspberry pi). The camera will almost always be facing down towards to ground, and the background will always be the same. I tried to make a haar cascade with positive samples from a kaggle dataset of 2K images of shoes, and a negative example with the background of office carpet textures. It did not work. I am looking into custom shape detection or something or that sort. If someone can help me find a way to do this besides object detection, I would apprciate it. the image attached is how the camera will be looking. it is fixed and the shoes will enter the frame occasionaly. I want to track the shoe and the position of the shoe.
unlikely, that you’ll get this idea to work, ever (broken plan)
the difference between 2 different shoe models might be larger than the diff between a shoe & carpet. (‘inner class variance’ is too high). also, both left & right shoe in the same model makes it worse.
hmm, got a link for us ?
a suitable neural network, suitably trained, should be able to spot “shoes” in such similar pictures, using contextual clues such as visible legs.
depending on how you’ll approach this, it could be trivial to train, or it could be a lot of preparation.