By transferring of prior knowledge from source domains and synthesizing the new knowledge extracted from the target domain, the performance of learning can be improved when there are insufficient training data in the target domain. In this paper we propose a new method to transfer a deformable part model (DPM) for object detection, using sharable filters from offline-trained auxiliary DPMs of similar categories and new filters learnt from the target training samples to improve the performance of the target object detector. A DPM consists of a collection of root and part filters. The filters of the auxiliary detectors capture the sharable appearance features and can be used as prior knowledge. The sharable filters are employed by the new detector with a coefficient reweighting algorithm to fit the target object much better. Meanwhile the target object still has some distinct local appearance features that the part filters in the auxiliary filter pool can not represent. Hence, new part filters will be learnt with the training samples of the target object and added to the filter pool as complementary. The final learnt model will be an assembly of transferred auxiliary filters and additional target filters. With a latent transfer learning algorithm, appropriate local features are extracted for the transfer of the auxiliary filters and the description of the distinct target filters. Our experiments demonstrate that the proposed strategy precedes some state-of-the-art methods.
High-performance pedestrian detection with good accuracy and fast speed is an important yet challenging task in computer vision. We design a novel feature named pair normalized channel feature (PNCF), which simultaneously combines and normalizes two channel features in image channels, achieving a highly discriminative power and computational efficiency. PNCF applies to both gradient channels and color channels so that shape and appearance information are described and integrated in the same feature. To efficiently explore the formidably large PNCF feature space, we propose a statistics-based feature learning method to select a small number of potentially discriminative candidate features, which are fed into the boosting algorithm. In addition, channel compression and a hybrid pyramid are employed to speed up the multiscale detection. Experiments illustrate the effectiveness of PNCF and its learning method. Our proposed detector outperforms the state-of-the-art on several benchmark datasets in both detection accuracy and efficiency.