Deep Neural Networks Approaches for Person and Vehicle Re-Identification.

English

The task of Re-Identification (re-id) is to retrieve any entity's images from a gallery set of multiple non-overlapping cameras by a given probe image. Active re-id research is composed of two entities named as person re-identification and vehicle re-identification. Person re-id is a very challenging task due to the presence of illumination, appearance, background, viewpoint, domain and pose variations, lighting and occlusions in the persons images. Domain variations occur due to camera's field of view environment (indoor and outdoor cameras) and light intensity as there are many different cameras present in the surveillance networks. Pose variations take place due to the person's movement and position when captured in different cameras. These variations make it difficult to match images of the same persons. To overcome this issue, we adopt Generative Adversarial Network (GAN) based approach to generate images in multiple domains and poses. Another solution for such problem is proposed by generating images from one camera domain to all other camera domains present in the environment and then merge these generated samples with the original data to enhance the method's performance. The proposed mechanisms magnify the matching between two images of the same person. The cameras used in video surveillance systems usually generate low resolution degraded images. The neural networks fail to learn various salient features in the existence of noise and other degradations in the data. The general architectures of neural networks learn features through neighbourhood similarities and hence forgot the similar patches which are discriminative for a specific person. The ignorance of such long-range similarities (dependencies) in the learned features halts the performance of neural networks. We introduce attention mechanisms in the existing neural networks to get rid of such limitations. Attention is used to include non local and distant computations in the local receptive field of the convolution operations in networks. We propose multiple designs and solutions by the addition of a couple of attention mechanism which improve the performance of the network in different aspects. Another active research line is cross resolution person re-id in which query and gallery images are of different resolutions and hence reduce the performance of the existing person re-id models. We propose a distillation process based on resolution features in addition with channel attention to tackle this problem. Vehicle re-id is different from person re-id in the aspects of orientations present and the existence of multiple vehicles of a same model in the data. Vehicles have limited number of colors and models so require more discriminative features to represent a specific vehicle. We propose an oriented splitting of the features to learn local features along with global features to create a strong descriptor for each vehicle.