Learning to Correct for Bad Camera Settings in Large Scale Plant Monitoring

June 18
Figure 1: Each subfigure shows the original image over the color-corrected image, and the masks derived from each.

Zongyang presented his research and implementation on an algorithm to correct image exposure at the CVPR 2019 workshop on "Computer Vision Problems in Plant Phenotyping".

Abstract: In large scale, automated image capture systems,incorrect camera settings can lead to images that are completelyuseless or which break assumptions made by image analysisalgorithms, but there may also be sufficient data to learn to automatically correct bad data. We consider the specific problem of over- and under-saturated images in large scale plant growthmonitoring, propose a generative approach that addresses thesesaturation issues for the calculation of plant canopy cover, and suggest future areas of research in this problem domain.

 

The source code for this algorithm is released under the BSD 3-Clause Open Source License and can be found in the TERRA REF Stereo RGB repository on GitHub.