This paper introduces a simple yet effective retrieval framework for object retrieval and localization. Our method is based on min-Hash method using compositional structure-preserved object representation. Compared to the traditional hash-based Content-based image retrieval (CBIR) system which always suffers from low recall due to insufficient discriminability of image representations, our method contributes in the following three terms: firstly, a new image feature, namely Pair of Geometric Coupled Words (PGCW), is presented to impose spatial context into visual words and generate very discriminative hash functions. Secondly, we select a batch of hashing functions by learning from a number of supervised retrievals. The sketches are then generated by selecting the hashing functions from the constructed object model. Finally, in the step of hash sketches matching, we introduce an auxiliary offset space, in which the object localization can be estimated by clustering. Our approach valids on popular public image databases and outperforms state-of-the-art methods.