Dp 507 en
Evaluation and Improvements of Image Interest Region Detectors and Descriptors
Author: Karel Lenc
A reliable and informative performance evaluation of local feature detectors and descriptors is a difficult task that needs to take into account many applications and desired properties of the local features.The main contribution of this work is the extension of the VLBenchmarks project which intends to collect major evaluation protocols of local feature detectors and descriptors.We propose a new benchmark which evaluates local feature detectors in the image retrieval tasks and simple epipolar criterion for testing detectors and descriptors in the wide baseline stereo problems. Using the extended benchmarks we investigate several parameters of the local feature detection algorithms.We propose a new algorithm for building a scale space pyramid which significantly improves the detector repeatability in the case of apriori knowledge of the nominal Gaussian blur in the input image. On the image retrieval tasks, we show that features with a small value of the response function improve the performance more than features with small scale, contrary to the observations in the geometry precision benchmarks. By altering the computation of the SIFT descriptor, we show that it is not necessary to weight the patch gradient magnitudes when input images are similarly oriented and that for blob-like features increasing the measurement region improves the performance.
Finally we propose an improvement of emulated detectors that allows finding new image features with better geometric precision. We have also improved the classification time of the emulated detectors and achieved higher performance than the handcrafted OpenSURF detector implementation.