currently outperforms more complex approaches, such as
that proposed by Desai et al. 8
We evaluated different aspects of our system on the lon-ger-established PASCAL VOC 2007 dataset. Figure 6 summarizes results of different models for the person category.
We trained models with 1 and 3 components, with and
without parts, and forcing mirror symmetry in each component or allowing for asymmetric models. We see that
the use of parts can significantly improve the detection
accuracy. Mixture models are also very important in the
figure 5. Examples of high-scoring detections on the PascaL 2007 dataset. the red-framed images (last two in each row) illustrate false
positives for each category. many false positives (such as for person and cat) are due to the stringent bounding box overlap criteria.