Cascaded SR-GAN for Scale-Adaptive Low Resolution Person Re-identification

Abstract

Person re-identification (REID) is an important task in video surveillance and forensics applications. Most of previous approaches are based on a key assumption that all person images have uniform and sufficiently high resolutions. Actually, various low-resolutions and scale mismatching always exist in open world REID. We name this kind of problem as Scale-Adaptive Low Resolution Person Re-identification (SALR-REID). The most intuitive way to address this problem is to increase various low-resolutions (not only low, but also with different scales) to a uniform high-resolution. SR-GAN is one of the most competitive image superresolution deep networks, designed with a fixed upscaling factor. However, it is still not suitable for SALR-REID task, which requires a network not only synthesizing high-resolution images with different upscaling factors, but also extracting discriminative image feature for judging person's identity. (1) To promote the ability of scale-adaptive upscaling, we cascade multiple SRGANs in series. (2) To supplement the ability of image feature representation, we plug-in a reidentification network. With a unified formulation, a Cascaded Super-Resolution GAN (CSR-GAN) framework is proposed. Extensive evaluations on two simulated datasets and one public dataset demonstrate the advantages of our method over related state-of-the-art methods.

HSME: Hypersphere Manifold Embedding for Visible Thermal Person Re-Identification

“…A large number of algorithms for Re-ID problem have been proposed, such as (Zheng, Yang, and Hauptmann 2016) . Recent researches mainly focus on visible pedestrian images (Zheng et al 2015) (Wang et al 2018b) (Wang et al 2016), i.e., both query images and gallery images are captured by visible camera. However, in night time or dark environment, visible images become uninformative.…”

Multi Scale-Adaptive Super-Resolution Person Re-Identification Using GAN

“…However, regardless of various LR and SV, most of these approaches are based on the premise that both query and gallery images are of the same or sufficiently high resolutions. This assumption may not hold in real-world re-identification problems as there happen to be many variations in the resolution of the person images under surveillance network [16]. These variations occur due to the natural changes in the distance between person and cameras, e.g., pedestrian movement relative to the camera.…”

“…To cope with this problem, an adequate strategy is to use an image super-resolution (SR) method to improve the resolution of LR query images to minimize the distribution discrepancy with HR gallery images. Several methods [16], [18]- [24] have been introduced to address the LR PREID problem. However, there are some common drawbacks in these methods: (1) Instead of trying to recover the misplaced discriminative appearance information, they perform a transformation of the cross resolution representation in pre-defined feature space [20]- [22], [24], [25].…”

“…Naturally, image (SR) would provide an effective way to minimize the dilemma of resolution discrepancy because of its ability to produce high-frequency details [26]. (3) Utilization of multibranch super-resolution networks to generate high-resolution person images like CSR-GAN, employed three consecutive GANs unit for image SR [16], [19]. It makes the overall network complex, requires more memory and power.…”

Residual Attribute Attention Network for Face Image Super-Resolution

“…Face super-resolution (SR), also known as face hallucination, is the process of recovering a high-resolution (HR) face from an input low-resolution (LR) face image. Face SR can be used as an important means of image preprocessing and widely used in various fields related to face images, e.g., face parsing (Li 2017), identity recognition (Taigman 2014;Wang, Ye, and Yang 2018;Wang, Hu, and Yu 2016), and face alignment (Jourabloo, Ye, and Liu 2017), where highfrequency face details are desired.…”