Concetti Chiave
Using HDR images as input to feature point detection and description algorithms improves performance compared to using LDR images, especially in scenes with extreme lighting conditions.
Sintesi
The study presents a systematic review of image feature point detection and description algorithms that use HDR images as input. The authors developed a library called CP HDR that implements the Harris corner detector, SIFT detector and descriptor, and two modifications of those algorithms specialized in HDR images, called SIFT for HDR (SfHDR) and Harris for HDR (HfHDR).
The key highlights and insights from the study are:
Most feature point detection and description algorithms are designed for low dynamic range (LDR) images, which can have issues in scenes with extreme light conditions due to under and overexposed areas. High dynamic range (HDR) images can be used to overcome these problems.
The authors conducted a systematic review to understand the state-of-the-art and list the datasets, algorithms, and metrics used in the literature. They found that most studies use tone mapping (TM) algorithms to transform HDR into LDR images before applying feature point extraction.
The CP HDR library can receive both LDR and HDR images as input to detection and description algorithms. The authors compared the performance of the algorithms when using LDR and HDR images as input.
Using uniformity, repeatability rate, mean average precision, and matching rate metrics, the results show that using HDR images as input to detection algorithms improves performance, and that SfHDR and HfHDR enhance feature point description.
The use of the coefficient of variation (CV) filter and logarithmic transformation in the HfHDR and SfHDR detectors helps improve feature point distribution in areas with different lighting conditions when using HDR images.
Statistiche
The number of feature points detected in the brightest, intermediate, and darkest areas of the image is more evenly distributed when using HDR images as input compared to LDR images.
The mean average precision and matching rate are higher when using HDR images as input to the feature point description algorithms.
Citazioni
"Using HDR images as input to detector and descriptor algorithms requires changing these algorithms to support the dynamic range of HDR images correctly."
"The use of the coefficient of variation (CV) filter and logarithmic transformation in the HfHDR and SfHDR detectors helps improve feature point distribution in areas with different lighting conditions when using HDR images."