Object Detection in Natural Images

This thesis work is aiming for detecting and localizing different classes of objects in natural images. For this, we use an algorithm that first detects objects using Implicit Shape model (ISM), then, with patch based class segmentation, verifies the hypothesis. This algorithm is able to achieve a high detection rate on both rigid and articulated objects. The thesis also proposes a novel multi-class ISM based on training a joint codebook which benefits from sharing features. Experiments verify that multi-class ISM can reach the same level of detection accuracy as the original ISM while being considerably more efficient. To further improve the computation efficiency, Information Gain(IG) and Document Frequency criteria are combined to select codebook entries. Evaluation shows that decreasing the codebook size by 85% doest not affect the detection accuracy. To further improve the performance, saliency detection can be applied on test images on certain datasets. This is also shown by the experiments.