Real-time multi-image vignetting and exposure correction for image stitching

Seamless image stitching depends not only on the accurate alignments of camera images, but also on the compensation of illumination inconsistencies. Even if two images are aligned perfectly, the seam is still visible if the images have a distinct vignetting or different exposure. Image stitching is used to expand the field of view, but a visible seam can lead to significant errors in subsequent visual perception tasks. As a result, we present a straightforward and accurate method for vignetting and exposure correction for stitched images. Firstly, we estimate the camera response function that maps irradiance to intensity. Then, the vignetting model is determined, which is applied to the irradiance images. After that, the exposure of the stitched images is corrected with the irradiance values at the seam. Finally, the irradiance is converted back into intensity using the camera response function. Our approach is evaluated using data recorded by our experimental vehicle and the public nuScenes dataset. Thereby, we test the performance of our method using the IoU of the histograms as well as the mean absolute error of the intensity values in the overlapping image regions. Furthermore, we demonstrate the real-time capability of our approach.