We are using a template descriptor on the image in order to try and find the object. However, we have a sparse 3D point clouds of the world that is not used at all when looking for the object in the images.Considering...We are using a template descriptor on the image in order to try and find the object. However, we have a sparse 3D point clouds of the world that is not used at all when looking for the object in the images.Considering there are many false alarms during the detection, we are interested in exploring how to combine the detections on the image with the 3D point clouds in order to reject some detection outliers. In this experiment we use semi-direct-monocular visual odometry(SVO) to provide 3D points coordinates and camera poses to project 3D points to 2D image coordinates. By un-projecting points in the tracking on the selection tree(TST) detection box back to 3D space, we can use 3D Gaussian ellipsoid fitting to determine object scales. By ruling out different scales of detected objects, we can reject most of the detection outliers of the object.展开更多
文摘We are using a template descriptor on the image in order to try and find the object. However, we have a sparse 3D point clouds of the world that is not used at all when looking for the object in the images.Considering there are many false alarms during the detection, we are interested in exploring how to combine the detections on the image with the 3D point clouds in order to reject some detection outliers. In this experiment we use semi-direct-monocular visual odometry(SVO) to provide 3D points coordinates and camera poses to project 3D points to 2D image coordinates. By un-projecting points in the tracking on the selection tree(TST) detection box back to 3D space, we can use 3D Gaussian ellipsoid fitting to determine object scales. By ruling out different scales of detected objects, we can reject most of the detection outliers of the object.
