Image object retrieval aims to retrieve database images in which image object query might only cover a small region. Object retrieval over a large-scale database is challenging because of the need to respond user query efficiently; existing methods suffer from low recall rate since target images differ visually from the query due to commonly observed occlusions, scaling, lighting variations, etc. Therefore, the success of an image object retrieval system hinges on the following issues: 1) finding the occurrences of specific objects in the image collections; 2) responding to queries over the million-scale collections in less than few seconds; 3) improving recall rate. To resolve these issues, we propose efficient query evaluation with pseudo-objects in the adapted inverted indexing to achieve time efficiency. In addition, we propose to augment each image with auxiliary visual features from image (textual and visual) clusters by informative feature selection and propagation, automatically mined offline in a distributed platform (i.e., MapReduce). Experiments show that the proposed framework takes 121ms for a query in the one-million collection; it achieves significant improvement in retrieval accuracy (e.g., up to 98.8% relative improvement) over the prior methods.