The Challenges of Image Classification
Before Convolutional Neural Networks came around, inferring what your images was a manual process of extracting features and piping them into a classifier. This project explores those classic techniques and compares three feature representations of images.
- Compared histogram oriented gradients (HOG) and bag of visual words (BoW) using SIFT descriptors to raw pixel data using a LinearSVM.
- Built BoW using OpenCV's implementation of SIFT and scikit-learn's KMeans.
The accuracy improved the more I abstracted the image, where SIFT created the largest accuracy, but I found that there was a problem with some classes not being guessed at all.
- Mini-Batch KMeans