Research on Pseudo-Label Improvement Based on Probabilistic Uncertainty

Abstract The pedestrian re-identification task explored in this study has strong application scenarios in real life. Generally, in actual scenarios, the pedestrian samples taken by the camera are unlabeled, so pseudo-labeling is used. However, low-quality image feature extraction and imperfect clustering will produce low-quality and incorrect clustering results, which will have a negative impact on the re-identification performance. This article proposes a dual extraction method (PLR). The core idea of this study is to explore the distribution differences and similarities between two data sets through the pseudo-label method, so that the model can learn the distribution in one data set and apply it to another data set and accurately identify the target person to be found. Therefore, this article proposes a dual refinement method based on probability uncertainty. PLR first removes unqualified clustering results through probability uncertainty, and then continues to refine the remaining ones. Experiments have shown that compared to newer methods, PLR is approximately 1% higher in most tasks.