I was working with the feature_tracker_nvxcu demo in VisionWorks 1.6 on my Jetson TX2, and I am interested in getting more data from the keypoints than just the x and y locations. So, I modified the feature_tracker_nvxcu.cpp file replacing all instances of nvxcu_point2f_t with nvxcu_keypointf_t and all instances of NVXCU_TYPE_POINT2F with NVXCU_TYPE_KEYPOINTF. This builds just fine, but the demo doesn’t run.
According to the visionworks cuda documentation, all of the functions used in the demo support both types of keypoint, but it crashes whenever I try to run it. I’d appreciate it if Nvidia could investigate this to see whether this problem is reproducible. I’m interested in getting the feature strength (and possibly the error, orientation, and scale) that is included in the keypoint struct (but which is absent from point2f). Thanks!