High-Resolution and Low-Resolution Video Person Re-Identification: A Benchmark

Abstract

Person re-identification has recently attracted increasing interest in the computer vision and safety-critical applications. In practice, due to poor quality of cameras or long distance away from person, the captured pedestrian videos usually suffer from low resolution, which will result in the loss of useful information contained in videos and make person re-identification between low-resolution (LR) and high-resolution (HR) videos (PRLHV) be a challenging task. However, the problem of PRLHV has not been well studied. In this paper, we propose a semi-coupled mapping based set-to-set distance learning (SMDL) approach for PRLHV. Specifically, by regarding each video as a set of features extracted from several walking cycles, we learn a discriminative set-to-set distance metric to enhance the separability between videos from different persons. To decrease the influence of low resolution on the distance learning, we design a clustering-based semi-coupled mapping term for our approach, which can reduce the variation between features of low-resolution and high-resolution videos by a semi-coupled mapping matrix. Since there exists no low-resolution video pedestrian re-identification dataset under real-world scenario up to now, we contribute a benchmark dataset for PRLHV, named high-resolution and low-resolution video person reidentification dataset (HLVID). Although this dataset is challenging and difficult for person re-identification, it is a useful attempt for further studies on low-resolution video-based pedestrian re-identification under the real-world scene. The extensive experiments on the newly collected video dataset demonstrate that our approach performs better than the state-of-the-art person re-identification methods in the PRLHV task. INDEX TERMS Distance learning, low-resolution video-based pedestrian dataset, semi-coupled mapping, person re-identification.

Part-Based Enhanced Super Resolution Network for Low-Resolution Person Re-Identification

“…Traditional image based person REID methods assume that all images are in high resolution, but a large variation in cross-view resolutions might occur in real applications. Various REID models have proposed to focus on this low resolution task, including [11], [27]- [29]. A common solution is to reduce the impact of different resolutions on identification task by learning the cross-mapping between LR and HR images.…”

“…The approach [11] is instantiated by designing a hybrid deep convolutional neural network for improving low resolution re-id performance and utilizing an adaptive fusion algorithm for accommodating multi-resolution LR images. To reduce the influence of low resolution on the distance learning, Ma et al [27] designed a clustering-based semi-coupled mapping term, which can reduce the variation between features of LR and HR videos by a semi-coupled mapping matrix. Another popular solution is to elevate LR images to a uniform HR [10], [11], [30], followed by identification tasks.…”

Beyond Intra-modality: A Survey of Heterogeneous Person Re-identification

“…Chen et al [55] proposed a network architecture of resolution adaptation and re-identification network to solve He-ReID LR problem. Ma et al [56] extended their focus of person re-identification to the low-resolution and high-resolution videos matching, and propose a semicoupled mapping based set-to-set distance learning method. Wang et al [53] first cascaded multiple SR-GANs in series to promote the ability of scale-adaptive upscaling, then plugged-in a re-identification network to supplement the ability of image feature representation.…”

Target Field of View Prediction Using Artificial Pheromones for People Reidentification

“…The joint discriminant optimal model on feedback top ranks matching pairs will enhance the discrimination of matching pairs similarity. For the same problem of overfitting, in [39], they proposed a semi coupled mapping-based set-to-set distance learning (SMDL) approach.…”