Runtime Model Recommendation for Exemplar-based Object Detection

We present an approach for object instance detection that uses model recommendation to predict a subset of relevant exemplar models for object detection based on an testing image at runtime. An initial subset of randomly selected exemplar models, the probe set, is first applied to the testing image, and its responses are used, in conjunction with a rating matrix, to predict the responses of all the exemplar models. The subset of exemplar models predicted to score the highest is then applied to the testing image to generate the final detections. This method enables scaling up the number of exemplar models to capture large object appearance variability, while maintaining computational efficiency. We present a novel max-selection scheme that allows us to build the rating matrix in a weakly-supervised fashion, allowing us to leverage large amounts of data easily. In addition to computational efficiency, we present experimental results which demonstrate that this model recommendation approach can outperform a baseline in which all the exemplar models are evaluated on the testing image.