Computer Vision News - September 2016

Computer Vision News Computer Vision News Research 31 Research on Convolutional Neural Networks . The framework is trained in an effective manner and outperforms the state-of-the-art methods on a number of benchmark datasets, without the need for retraining. Method: The LIFT pipeline The Detector (DET), the Orientation Estimator (ORI) and the Descriptor (DESC) are all CNN-based, coupled with an end-to-end differentiability framework. For training, the LIFT framework uses a four-branch Siamese architecture (see figure below), where each branch contain three CNNs marked by DET, ORI and DESC. More details on these three networks follow. The Siamese architecture: given an input image patch P, the Detector (DET) provides a score map S, which feeds into the softargmax and returns the location x of one feature point. Then, a smaller patch p centered on x is extracted along with the Spatial Transformer Crop. Next, the patch is used as the input to the Orientation Estimator (ORI), to predict the patch orientation and rotate it according (Rot). Finally, this patch is fed into the Descriptor network, which computes a feature vector d.