Beyond neighbourhood-preserving transformations for quantization-based unsupervised hashing

• A new type of non-rigid transformation is presented for quantization based hashing. The proposed corrupts the neighborhood structure data in favor quantization. method reduces dimensionality and error simultaneously. Unlike other methods, we employ both rigid transformations to compress data. An efficient sequential scheme update optimization variables with low training time. effective unsupervised hashing algorithm leads compact binary codes preserving as much possible. One most established schemes reduce then find a (neighborhood-preserving) that error. Although employing effective, may not loss ultimate limits. As well, reducing two separate steps seems be sub-optimal. Motivated by these shortcomings, propose We relax orthogonality constraint on projection PCA-formulation regularize this term. show matrix rotation contribute towards minimizing but different ways. scalable nested coordinate descent approach optimize mixed-integer problem. evaluate five public benchmark datasets providing almost half million images. Comparative results indicate mostly outperforms state-of-art linear methods competes end-to-end deep solutions.