Abstract
In this work, we propose an ultra-fast unsupervised saliency detection algorithm based on QR matrix factorization. The algorithm works by dividing an image into blocks and computing the QR factors in each local neighborhood where salient patches correspond to columns in R with largest L-0 norm. Experimental results on the MSRA10K salient object database show that the proposed algorithm is competitive with state-of-the-art real-time saliency detection algorithms in terms of AUC and execution time. Additionally, we study the various parameters that control the performance of the proposed algorithm such as threshold values and processing-block sizes and their effect on the overall performance.