Paper discussion blog for CS 2951t at Brown University. Instructor: Genevieve Patterson
In Three-Dimensional Object Detection and Layout Prediction using Clouds of Oriented Gradients, Ren and Sudderth propose a new feature representation along with algorithms to improve three-dimensional object detection and layout prediction. They generalize the HOG image feature to their three-dimensional Cloud of Oriented Gradients (COG) feature. They apply a particular kind of SVM to classify and give bounding polyhedra to common household objects such as tables, sofas, toilets and chairs. They also build and train a graphical model that models the context of the scene, and obtain state of the art performance using an IOU metric on all categories tested. Discussion: What steps could be taken to get this system to perform faster than ~10-30 minutes per object category towards say real-time video? How much speedup would a GPU give?