AS-RIG: Adaptive selection of reconstructed input by generator or interpolation for person re-identification in cross-modality visible and thermal images

Abstract

Multimodal camera-based person re-identification (ReID) is important in the field of intelligent surveillance. Thermal cameras can solve the problem in that visible-light cameras cannot acquire the valid feature information of a person under poor illumination conditions. However, thermal cameras usually have lower frame resolution than visible-light cameras. To overcome this problem, we propose an adaptive selection of reconstructed input by generator or interpolation (AS-RIG) method, which can adaptively select the generative adversarial network (GAN), or an interpolation method (bi-linear or bi-cubic). AS-RIG automatically selects a resolution-model using the mean-squared error (MSE), feature distance (FD), and structural similarity (SSIM). To verify the performance of our proposed method, two open databases are used: the DBPerson-Recog-DB1 and Sun Yat-set University multiple modality Re-ID (SYSU-MM01). Infrared frames from both databases are resized to be smaller than the original ones for experimentation. Experimental results show that our generator outperforms traditional interpolation methods. In addition, the person ReID experimental results demonstrate that AS-RIG outperforms non-adaptive selection methods and state-of-the-art methods.