Center for Machine Perception
● Proposed a new local feature descriptor that improves matching scores on all standard benchmarks.
● Introduced an effective method for dataset reduction - showed that reducing the size 10 times improves results.
● Developed a method allowing training of a descriptor on big data (10TB AMOS dataset), improving results on a standard benchmark.
● Introduced a new benchmark for testing of robustness of a descriptor to illumination changes.
● Presented a method for compression of network outputs that also improves performance.