A method for scene frame matching for automatic content recognition. The method includes receiving fingerprints indicative of frames of a media stream. The method also includes obtaining scene centroids of corresponding scenes of the media stream. Each scene includes fingerprints ranging from an initial frame having a corresponding initial fingerprint and a last frame having a corresponding last fingerprint. Each scene centroid represents a mean of fingerprints corresponding to a respective scene. The mean of fingerprints is determined by an average of pixel values corresponding to the fingerprints of the respective scene. For each scene centroid, the method includes determining a distance between the respective scene centroid and a target fingerprint of a corresponding target frame of the media stream and determining a target scene of the target frame as being the respective scene of the scene centroid having a shortest distance to the target fingerprint of the target frame.