Omni-Scale Feature Learning for Person Re-Identification

Abstract

As an instance-level recognition problem, person reidentification (ReID) relies on discriminative features, which not only capture different spatial scales but also encapsulate an arbitrary combination of multiple scales. We callse features of both homogeneous and heterogeneous scales omni-scale features. In this paper, a novel deep ReID CNN is designed, termed Omni-Scale Network (OSNet), for omni-scale feature learning. This is achieved by designing a residual block composed of multiple convolutional feature streams, each detecting features at a certain scale. Importantly, a novel unified aggregation gate is introduced to dynamically fuse multiscale features with input-dependent channel-wise weights. To efficiently learn spatial-channel correlations and avoid overfitting, the building block uses both pointwise and depthwise convolutions. By stacking such blocks layerby-layer, our OSNet is extremely lightweight and can be trained from scratch on existing ReID benchmarks. Despite its small model size, our OSNet achieves state-ofthe-art performance on six person-ReID datasets. Code and models are available at: https://github.com/ KaiyangZhou/deep-person-reid.

Discriminative Spatial Feature Learning for Person Re-Identification

“…We compare the proposed method with 33 recent published works including (1) global feature based methods which aims to learn the global feature from the feature map directly, including PAN [74], DMML [7], DCDS [1], VCFL [30], MVPM [41], LRDNN [79], RB [35], LITM [63], IANet [23], Sphere [14], BNNeck [32], OSNet [78], AANet [46], DG-Net [72], BDB [12], Circle [42], SFT [31], (2) part based methods including PCB+RPP [43], Local [57], HPM [16], CASN [71], AutoReID [34], MGN [49], BHP [20] and Pyramidal [68] which utilize the semantic parts or horizontal stripes to extract part-level feature, and (3) attention based methods including MHAN [3], CAMA [58], SONA [53], CAR [80], SCAL [6], ABD-Net [8], DAAF [10] and RGA [65]. These methods are categorized into 3 types based on different backbones: the ones which employ ResNet-50 directly, the ones which modify ResNet-50 by introducing additional branches, attention subnets or dilated convolution, and the others which don't use ResNet-50.…”

“…We mainly review the former which utilize deep learning to extract the feature. Holistic Features Based Methods Given a backbone C-NN such as ResNet-50 [21] or other network architectures [2,51,71,78], this type of methods learns discriminative holistic features from the feature map directly. Specifically, they aim to learn the features by improving loss functions [9,14,22,31,41,42,50,55,63], improving the training techniques [1,4,12,24,32,35,37,54], adding additional network modules [23,23,51,62], using extra semantic annotations [30,46,47,79] or generating more training samples [17,33,72,76,77].…”

An Effective Adversarial Attack on Person Re-Identification in Video Surveillance via Dispersion Reduction

“…By learning jointly on global and local features, it aims to address existing drawbacks. Xie et al [51] proposed PLR-OSNet, which introduces Part-level resolution (PLR) into Omni-Scale Network (OSNet) [52]. It has two branches including both global and local feature representations.…”

REVAMP2T: Real-Time Edge Video Analytics for Multicamera Privacy-Aware Pedestrian Tracking

“…The heart of pedestrian tracking is consistent reidentification (ReID) of those pedestrians throughout the frames of videos across multiple cameras. Similarly, on the re-identification side, recent methods leverage CNNs to extract unique features among persons [17][18][19][20][21][22][23][24]. The work in [25] learns the spatial and temporal behavior of objects by translating the feature map of the Region of Interest (RoI) into an adaptive body-action unit.…”