Object detection based on ensemble of exemplars

This paper proposes a detection approach for localizing the object of specific category in images. Based on the ensemble of exemplars, a per-exemplar classifier for each exemplar is learnt, which is simple but powerful to perform well in detecting visually similar objects. Meanwhile, considering the fact that the number of negatives is always considerably larger than that of positives, the method of hard-negatives mining is employed. In addition, using multiple instance learning, we also learn the per-category classifier with inputting the detections produced by per-exemplar classifier in the validation set. The better performance on object detection is boosted further with the contribution of co-occurrence matrices that encodes the relationship of each per-exemplar classifier. In the experiment section, we evaluate the performance of our approach on PASCAL VOC 2007 dataset and compare it to Tomasz´s Exemplar-SVM model directly. The experimental result demonstrates that our approach outperforms the Exemplar-SVM model in object detection with higher average precision.