keyword
Hamming embedding
Hamming embedding is an approximate nearest neighbor search technique used in image retrieval that maps high-dimensional local feature descriptors into compact binary signatures to improve matching accuracy. In visual bag-of-words representations, continuous image descriptors are partitioned into discrete visual words, which can introduce quantization errors by grouping dissimilar features into the same cluster. Hamming embedding addresses this by projecting the descriptor space assigned to each visual word into a lower-dimensional space and binarizing the resulting values using coordinate-wise thresholds. During retrieval, two descriptors are considered a valid match only if they share the same visual word and the Hamming distance between their binary signatures remains below a predefined threshold, effectively filtering out false positive matches while preserving high computational efficiency within inverted index structures.
1 item
