Rating is available when the video has been rented.
This feature is not available right now. Please try again later.
Published on Apr 5, 2016
Unoriented LATCH descriptor and matcher running on GTX 970M with CUDA. Symmetric matches are enforced, and ~3 billion comparisons are made per second. Open source: https://github.com/csp256/cudaLATCH (note, this version is outdated and an order of magnitude lower performance than current in both extraction and matching. Updates to come)
Roll, pitch, yaw, and translation direction are approximations only (camera is uncalibrated). Really the VO code is only present to put a load on the CPU for profiling reasons. Threshold refers to the threshold for the FAST detector, and is reported at 100 times actual value to be on the same scale as the number of keypoints and matches per frame. (Obviously a real implementation would need to make sure the keypoints were better distributed.)
Reported GPU run time is asynchronous: if the green line in the third plot window is below the red line, the GPU code is effectively instantaneous (only microseconds of overhead). Total overhead is actually about 43ms for all 4250 frames in this run.
The regular spikes in CPU run time are due to periodic plotting of the matches. CPU slow down during roll is due to lack of actual camera intrinsics and subsequent RANSAC failure from static inlier threshold in 5 point method.