Using the Deformable Part Model with Autoencoded Feature Descriptors for Object Detection

Future considerations would include ne tuning the latent SVM used to train the DPM, using stacked autoencoders to learn more complex feature representations, and optimize runtime of training algorithm to allow for larger training sets. Acknowledgements Special thanks to Professor Andrew Ng and Adam Coates for the advice they provided through the course of this project. Given its performance in recent years' PASCAL Visual Object Classes (VOC) Challenge, the Deformable Part Model (DPM) is widely regarded to be one of the state-of-the-art object detection and localization algorithms. The DPM as described by Felzenszwalb, et. al. uses Histogram of Oriented Gradients (HOG) descriptors as the underlying feature representation for an object. In this paper, we consider using features learned by a single-layered, sparse autoencoder as a substitute for HOG descriptors in the DPM. The rationale for this is that these learned feature descriptors may capture additional details present in the image that are not re ected in a human-designed set of features such as HOG. Using this more descriptive set of features may in turn yield a better object detector. Additionally, while it is true that deep learning algorithms such as the sparse autoencoder alone generally do not perform as well compared to other vision algorithms, it may be possible to integrate these learned feature descriptors into an existing image classi cation system such as the DPM. To further evaluate the performance of such an integration, we also consider the e ect of factors such as block normalization and colored features on the performance of our autoencoder-backed DPM. Finally, we examine the perfomance of the autoencoder-backed DPM across several di erent object classes from the PASCAL VOC Challenge.