Section: New Results

Improved Deformable Part Models for Object Detection

Paticipants: Iasonas Kokkinos, Stavros Tsogkas, Eduard Trulls, Pierre-Andre Savalle, George Papandreou.

In [30] and [36] we have worked on improving the classification accuracy of Deformable Part Models (DPMs) for object detection in two distinct manners. Firstly, in [30] we propose a technique to combine bottom-up segmentation, coming in the form of SLIC superpixels, with sliding window DPM detectors. The merit of our approach lies in ‘cleaning up’ the low- level features by exploiting the spatial support indicated by segmentation. - tion, for both the root and part filters of DPMs. We use these masks to construct enhanced, background- invariant features to train DPMs. We test our approach on the PASCAL VOC 2007, outperforming the standard DPM in 17 out of 20 classes, yielding an average increase of 1.7AP. Additionally, we demonstrate the robustness of this ap- proach, extending it to dense SIFT descriptors for large dis- placement optical flow.

Secondly, in [36] we have explored the potential of convolutional neural networks as feature extractors for detection with DPMs. In particular, we substitute the Histogram-of-Gradient features of DPMs with Convolutional Neural Network (CNN) features, and demonstrate that we thereby obtain a substantial boost in performance (+14.5 mAP) when compared to the baseline HOG-based models. Some more recent extensions to this work are included in [41] where we explore the potential of explicit scale and aspect ratio search in the context of sliding window detection with CNNs.