Feature point descriptors: infrared and visible spectra

Area of Science:

• Computer Vision
• Image Processing
• Infrared Imaging

Background:

• Feature point descriptors are crucial for image matching and analysis.
• Their performance in the long-wave infrared (LWIR) spectrum is less understood than in the visible spectrum.
• Understanding cross-spectral performance is vital for applications using multi-modal imaging.

Purpose of the Study:

• To evaluate the behavior of classical feature point descriptors in LWIR images.
• To compare their performance against results from the visible spectrum.
• To assess robustness against common image degradations.

Main Methods:

• Utilized a state-of-the-art evaluation framework.
• Tested robustness to rotation, scaling, blur, and additive noise.
• Employed a cross-spectral outdoor image dataset for experiments.

Main Results:

• Classical feature descriptors exhibit varied performance in LWIR compared to visible light.
• Robustness to image transformations and noise differs significantly between spectral bands.
• Specific descriptors show better or worse adaptation to LWIR characteristics.

Conclusions:

• The effectiveness of classical feature descriptors is spectrally dependent.
• LWIR imaging presents unique challenges for traditional feature detection and description.
• Results inform the selection of appropriate descriptors for cross-spectral computer vision tasks.