[Paper Review] An Effective Data Augmentation for Person Re-identification.

Yunpeng Gong, Zhiyong Zeng|arXiv (Cornell University)|Jan 21, 2021

Video Surveillance and Tracking Methods4 citations

TL;DR

This paper proposes a novel data augmentation method for person re-identification (ReID) that enhances model robustness to illumination variations by leveraging structural information from grayscale images. By combining Random Grayscale Transformation and Random Grayscale Patch Replacement, the method improves retrieval accuracy by up to 3.3% across multiple datasets, demonstrating significant performance gains over existing approaches.

ABSTRACT

In order to make full use of structural information of grayscale images and reduce adverse impact of illumination variation for person re-identification (ReID), an effective data augmentation method is proposed in this paper, which includes Random Grayscale Transformation, Random Grayscale Patch Replacement and their combination. It is discovered that structural information has a significant effect on the ReID model performance, and it is very important complementary to RGB images ReID. During ReID model training, on the one hand, we randomly selected a rectangular area in the RGB image and replace its color with the same rectangular area grayscale in corresponding grayscale image, thus we generate a training image with different grayscale areas; On the other hand, we convert an image into a grayscale image. These two methods will reduce the risk of overfitting the model due to illumination variations and make the model more robust to cross-camera. The experimental results show that our method achieves a performance improvement of up to 3.3%, achieving the highest retrieval accuracy currently on multiple datasets.

Motivation & Objective

• To improve person re-identification (ReID) performance by effectively utilizing structural information from grayscale images.
• To reduce the adverse impact of illumination variations on ReID models during training.
• To enhance model generalization and robustness across different cameras by introducing diverse grayscale augmentations.
• To achieve state-of-the-art performance on multiple benchmark ReID datasets through improved data augmentation.

Proposed method

• The method introduces Random Grayscale Transformation, which converts entire images to grayscale to expose models to monochromatic features.
• It applies Random Grayscale Patch Replacement, where randomly selected rectangular regions in RGB images are replaced with corresponding grayscale patches from the same image's grayscale version.
• The combination of both techniques generates training samples with mixed color and grayscale regions, increasing feature diversity.
• The approach is designed to reduce overfitting to specific illumination conditions by exposing the model to varied lighting representations.
• The augmentation is applied during training without modifying the inference pipeline, ensuring compatibility with standard ReID frameworks.

Experimental results

Research questions

• RQ1 How can structural information from grayscale images improve person ReID performance?
• RQ2 To what extent does data augmentation using grayscale transformations reduce sensitivity to illumination variation?
• RQ3 Can combining global and local grayscale augmentations lead to better generalization in cross-camera ReID?
• RQ4 What is the maximum performance gain achievable through this augmentation strategy on standard ReID benchmarks?

Key findings

• The proposed method achieves a performance improvement of up to 3.3% on multiple person re-identification datasets.
• The combination of Random Grayscale Transformation and Patch Replacement yields the highest retrieval accuracy reported to date on the evaluated benchmarks.
• Structural information from grayscale images significantly enhances model performance and serves as a valuable complement to RGB features.
• The method effectively reduces overfitting to illumination variations, improving robustness across different camera views.