Robust Person Re-Identification by Modelling Feature Uncertainty

Abstract

We aim to learn deep person re-identification (ReID) models that are robust against noisy training data. Two types of noise are prevalent in practice: (1) label noise caused by human annotator errors and (2) data outliers caused by person detector errors or occlusion. Both types of noise pose serious problems for training ReID models, yet have been largely ignored so far. In this paper, we propose a novel deep network termed DistributionNet for robust ReID. Instead of representing each person image as a feature vector, DistributionNet models it as a Gaussian distribution with its variance representing the uncertainty of the extracted features. A carefully designed loss is formulated in DistributionNet to unevenly allocate uncertainty across training samples. Consequently, noisy samples are assigned large variance/uncertainty, which effectively alleviates their negative impacts on model fitting. Extensive experiments demonstrate that our model is more effective than alternative noise-robust deep models.

Unsupervised Domain Adaptation for Person Re-Identification with Few and Unlabeled Target Data

“…Variances of this loss, also referred to as triplet loss, are adopted by the methods in [4,11,22,24,37]. Some works proposed novel loss functions [8,38]. Finally, combinations of losses are incorporated in [2,21,38,40].…”

“…Some works proposed novel loss functions [8,38]. Finally, combinations of losses are incorporated in [2,21,38,40].…”

“…Other methods directly integrate pose estimation [23,25,32] or dense part correspondence [40] for extracting part-aware features. In contrast to competing works, [21,38,45] do not adopt an ordinary backbone network, but propose their own. Finally, some recent works proposed modified architectures that better generalize to new domains [14,44].…”

Stable Median Centre Clustering for Unsupervised Domain Adaptation Person Re‐Identification

“…Since this method relies on clustering to assign the same pseudolabel to samples belonging to the same cluster and then optimizes the model with the pseudolabel as the supervision information [19], the credibility of the pseudolabel determines the performance of the model [20]. If we have a highly credible pseudolabel, the model can be adapted to the target domain data [21][22][23]. However, in the original dataset, there may be some samples with label noise, and the noisy samples interfere with the model training process because of their incorrect information in the feature space.…”

Data Covariance Learning in Aesthetic Attributes Assessment

“…These challenges make AAA more complicated. As pointed out in [17] modeling data uncertainty is beneficial in noisy data training [18] [19] [20] [21]. However, recent data uncertainty learning methods treat the modeled multivariate noise independent among components for simplicity, which is not an appropriate assumption for intertwined variables like photographic style.…”