Vision through distortions: Atmospheric Turbulence-and Clothing-invariant long-range recognition

Gabriel Bertocco, Fernanda A. Andaló, Terrance E. Boult, Anderson Rocha

Abstract: Long-range recognition is paramount in securitysensitive settings. It faces the hard task of retrieving images from a high-resolution gallery given a probe image affected by distortions due to atmospheric turbulence and different features, such as clothing. This work proposes a novel atmospheric turbulenceand clothing-invariant whole-body model to address the longrange recognition task. It leverages self-defined proxies across different acquisition ranges, a novel way to create diverse batches based on capturing condition and clothing, and a condition- and clothing-aware loss function. As most whole-body benchmarks have limited ranges, we employ the BRIAR dataset for training and evaluation. It comprises identities captured within 100 to $\mathbf{1, 0 0 0}$ meters from the camera in various poses, lighting conditions, and clothing variations. Quantitative and qualitative analysis show our model leads to distortion-invariant discriminative features across different recording capturing ranges. It also obtains competitive performance compared to the state-of-the-art benchmarks Market1501, MSMT17, and DeepChange.