What Image Classifiers Really See – Visualizing Bag-of-Visual Words Models

Bag-of-Visual-Words (BoVW) features which quantize and count local gradient distributions in images similar to counting words in texts have proven to be powerful image representations. In combination with supervised machine learning approaches, models for nearly every visual concept can be learned. BoVW feature extraction, however, is performed by cascading multiple stages of local feature detection and extraction, vector quantization and nearest neighbor assignment that makes interpretation of the obtained image features and thus the overall classification results very difficult. In this work, we present an approach for providing an intuitive heat map-like visualization of the influence each image pixel has on the overall classification result. We compare three different classifiers (AdaBoost, Random Forest and linear SVM) that were trained on the Caltech-101 benchmark dataset based on their individual classification performance and the generated model visualizations. The obtained visualizations not only allow for intuitive interpretation of the classification results but also help to identify sources of misclassification due to badly chosen training examples.