Compact descriptors of digital images are produced by detecting interest points representative of the digital images and selecting out of the interest points key points for producing e.g. local and global compact descriptors of the images. The digital images are decomposed into blocks by computing an energy (variance) for each said block and then subjecting the blocks to culling by rejecting those blocks having an energy failing to pass an energy threshold. The interest points are detected only in the blocks resulting from culling, and the key points for producing the compact descriptors are selected out of the interest points thus detected, possibly by using different selection thresholds for local and global compact descriptors, respectively. The number of key points for producing the compact descriptors may be varied e.g. by adaptively varying the number of the interest points detected in the blocks resulting from culling.