2016/05/26 by Edward Johns, Johns, Edward, Stefan Leutenegger +3 · 1 citation
Computer Science · Engineering · #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Image Processing Techniques and Applications #Robotics (cs.RO) #Robotics and Sensor-Based Localization
paper · pdf · doi:10.48550/arxiv.1605.08359
openalex publication_date 2016/05/26 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
A multi-view image sequence provides a much richer capacity for object\nrecognition than from a single image. However, most existing solutions to\nmulti-view recognition typically adopt hand-crafted, model-based geometric\nmethods, which do not readily embrace recent trends in deep learning. We\npropose to bring Convolutional Neural Networks to generic multi-view\nrecognition, by decomposing an image sequence into a set of image pairs,\nclassifying each pair independently, and then learning an object classifier by\nweighting the contribution of each pair. This allows for recognition over\narbitrary camera trajectories, without requiring explicit training over the\npotentially infinite number of camera paths and lengths. Building these\npairwise relationships then naturally extends to the next-best-view problem in\nan active recognition framework. To achieve this, we train a second\nConvolutional Neural Network to map directly from an observed image to next\nviewpoint. Finally, we incorporate this into a trajectory optimisation task,\nwhereby the best recognition confidence is sought for a given trajectory\nlength. We present state-of-the-art results in both guided and unguided\nmulti-view recognition on the ModelNet dataset, and show how our method can be\nused with depth images, greyscale images, or both.\n