Object Detection Using Keygraphs

We propose a new framework for object detection based on a generalization of the keypoint correspondence framework. This framework is based on replacing keypoints by keygraphs, i.e. isomorph directed graphs whose vertices are keypoints, in order to explore relative and structural information. Unlike similar works in the literature, we deal directly with graphs in the entire pipeline: we search for graph correspondences instead of searching for individual point correspondences and then building graph correspondences from them afterwards. We also estimate the pose from graph correspondences instead of falling back to point correspondences through a voting table. The contributions of this paper are the proposed framework and an implementation that properly handles its inherent issues of loss of locality and combinatorial explosion, showing its viability for real-time applications. In particular, we introduce the novel concept of keytuples to solve a running time issue. The accuracy of the implementation is shown by results of over 800 experiments with a well-known database of images. The speed is illustrated by real-time tracking with two different cameras in ordinary hardware.

A Keygraph Classification Framework for Real-Time Object Detection

In this paper, we propose a new approach for keypoint-based object detection. Traditional keypoint-based methods consist in classifying individual points and using pose estimation to discard misclassifications. Since a single point carries no relational features, such methods inherently restrict the usage of structural information to the pose estimation phase. Therefore, the classifier considers purely appearance-based feature vectors, thus requiring computationally expensive feature extraction or complex probabilistic modelling to achieve satisfactory robustness. In contrast, our approach consists in classifying graphs of keypoints, which incorporates structural information during the classification phase and allows the extraction of simpler feature vectors that are naturally robust. In the present work, 3-vertices graphs have been considered, though the methodology is general and larger order graphs may be adopted. Successful experimental results obtained for real-time object detection in video sequences are reported.