Triplet Loss with Channel Attention for Person Re-identification

Abstract

The triplet loss function has seen extensive use within person re-identification. Most works focus on either improving the mining algorithm or adding new terms to the loss function itself. Our work instead concentrates on two other core components of the triplet loss that have been under-researched. First, we improve the standard Euclidean distance with dynamic weights, which are selected based on the standard deviation of features across the batch. Second, we exploit channel attention via a squeeze and excitation unit in the backbone model to emphasise important features throughout all layers of the model. This ensures that the output feature vector is a better representation of the image, and is also more suitable to use within our dynamically weighted Euclidean distance function. We demonstrate that our alterations provide significant performance improvement across popular reidentification data sets, including almost 10% mAP improvement on the CUHK03 data set. The proposed model attains results competitive with many state-of-the-art person re-identification models.

Unifying Person and Vehicle Re-Identification

“…mAP rank-1 JLDE [36] 67.7 85.2 PAN [24] 69.3 86.7 SAN [57] 70.1 85.9 IC-TL [37] 70.1 86.6 ADV [38] 70.4 86.8 RE [60] 71.3 87.1 DaF [61] 72.4 82.3 DPFL [3] 73.1 88.9 SE+DWE [10] 74.2 88.0 MGCAM [62] 74.3 83.8 HA-CNN [21] 75 Method rank-1 rank-5 XVGAN [6] 60.20 77.03 NuFACT [4] 76.76 92.79 VAMI [63] 77.03 90.83 PROVID [4] 81.56 95.11 HA-CNN [21] 83.00 92.41 TriNet [1] 83 VeRi: The VeRi data set differs from other re-ID data sets as it maps temporally close images in the gallery onto tracks. The re-identification is computed from the query to the entire track (image-to-track) rather than just to gallery images (image-to-image).…”

“…For example, one may improve person re-ID performance by generating a better feature representation of pedestrians, e.g. by introducing squeeze and excitation modules [10]. This tells us nothing about the framework's actual ability to re-identify an object, despite the accuracy increasing.…”

“…Research focus in re-identification has shifted towards deep learning methods, which are routinely used to obtain state-of-the-art results over a wide variety of challenges in computer vision and machine learning. Typically, two types of CNN model have been employed to solve the person re-ID task: the classification model that is used across a broad spectrum of computer vision problems [16] and, more commonly for re-ID, the Siamese model which takes multiple images as input, such as pairs [17], [18], triplets [10], [19], and quadruplets [2].…”