BoVW

Bag of Visual Words (BoVW) is a classic image representation used in computer vision for image classification and object recognition. It adapts the bag-of-words idea from text to images by treating local image features as visual words and describing an image by the distribution of these words.

A typical BoVW pipeline consists of four stages. First, local features are detected and described from images using methods such as SIFT, SURF, or ORB. Second, a codebook or vocabulary is created by clustering a large collection of descriptors from the training set, usually with k-means, resulting in a fixed number of visual words. Third, each image is encoded by assigning its descriptors to the nearest visual word (hard assignment) or by probabilistic/soft assignment, and a histogram is formed that counts the occurrences of each word. Fourth, the histograms are fed into a classifier such as a support vector machine or logistic regression.

Extensions and variants include soft assignment to mitigate quantization errors, TF-IDF weighting to downweight ubiquitous words, and spatial information integration through Spatial Pyramid Matching, which layers histograms at multiple spatial scales. Limitations include the neglect of spatial layout within a single image, dependence on the chosen descriptors and codebook size, and computational costs for large vocabularies. BoVW served as a strong baseline before deep learning dominated many vision tasks, but it remains a reference point and can be combined with spatial pooling or kernel methods for enhanced performance.

Applications include image and object classification, scene recognition, and content-based image retrieval.