Development of a Pseudo-HDR Preprocessing Pipeline for the Improvement of LDR Saliency Algorithms

Author(s):

Darius Jefferson II and Andre Harrison

Abstract:

High dynamic range (HDR) imagery has an expanded luminance range capable of more accurately capturing the full range of luminance in the real world. However, most computer vision (CV) models cannot easily perceive the very dark or very bright areas in real-world HDR environments. This is because these models are typically trained on low dynamic range (LDR) imagery, which precondition CV models to expect a much smaller ratio between the brightest and darkest pixel in an image. Datasets of annotated HDR imagery are essential to training HDR-capable CV models for operation in real-world environments.

Saliency models are a type of CV model that would benefit from being retrained on HDR imagery. HDR imagery provides more information to the model for more accurate predictions. From a U.S. Army perspective, these retrained models could be used on autonomous systems in the battlefield to better inform Soldiers about areas in their environment that contain potential threats. However, very few HDR saliency datasets exist and none are large enough to train a model from scratch. In this report, we describe the development of a pipeline that would convert LDR image datasets into pseudo-HDR to train an HDR-adapted saliency model.