FlipReID: Closing the Gap between Training and Inference in Person Re-Identification

TL;DR

FlipReID introduces a training method that aligns training and inference features in person re-identification, leading to improved performance by minimizing the gap caused by test-time augmentation practices.

Contribution

The paper proposes the FlipReID structure with a flipping loss to unify training and inference features, enhancing re-identification accuracy.

Findings

Achieves state-of-the-art results on MSMT17 dataset.

Consistent performance improvements across multiple benchmarks.

Effectively reduces the gap between training and inference features.

Abstract

Since neural networks are data-hungry, incorporating data augmentation in training is a widely adopted technique that enlarges datasets and improves generalization. On the other hand, aggregating predictions of multiple augmented samples (i.e., test-time augmentation) could boost performance even further. In the context of person re-identification models, it is common practice to extract embeddings for both the original images and their horizontally flipped variants. The final representation is the mean of the aforementioned feature vectors. However, such scheme results in a gap between training and inference, i.e., the mean feature vectors calculated in inference are not part of the training pipeline. In this study, we devise the FlipReID structure with the flipping loss to address this issue. More specifically, models using the FlipReID structure are trained on the original images and the flipped images simultaneously, and incorporating the flipping loss minimizes the mean squared error between feature vectors of corresponding image pairs. Extensive experiments show that our method brings consistent improvements. In particular, we set a new record for MSMT17 which is the largest person re-identification dataset. The source code is available at https://github.com/nixingyang/FlipReID.