Re-identification Robustness Over Time - The Case for Synthetic Training Data

This study proposes the use of synthetic data to improve the robustness of training and testing of re-identification models in the context of logistics. In particular, this work considers image generation and re-identification of pallet blocks. At first, synthetic data is generated by multiple, small shifts through the latent feature space of a Generative Adversarial Network. As the shifts are performed in specialized latent feature space directions, the resulting images depict the same image identity while offering slightly modified visual features, like thickened branding patterns or different lighting. To guarantee that the generated images are appropriate for the training and testing of re-identification models, the image identity may not be destroyed through the modifications. Therefore, a threshold based on the Multi-Scale Structural Similarity Index is applied. The artificial data is then used alongside real data to train re-identification models with the goal of improving robustness while reducing the need for real data. To create common ground for pallet block re-identification, the dataset pallet-block-98382_3270 is released alongside this study. The dataset consists of 280,763 images of 98,382 real wooden Euro-pallet blocks and an additional 42,510 artificially generated images of 3,270 pallet block identities which were selected as suitable for the training and testing of re-identification models. Additionally, the dataset contains 143,000 unfiltered artificially generated images. The use of additional artificial data during training increases the mean Average Precision by up to 1.3%, indicating that real data can be replaced or supplemented by synthetic data. Even when not increasing the mean Average Precision, the use of synthetic data can be recommended as it increases the re-identification models' robustness across different testing datasets. These insights are crucial in industrial contexts, where data acquisition is costly and limited.

I. Introduction

Automated tracking of materials and goods in warehouses is necessary to ensure faultless and efficient operations. Logistics carriers such as wooden pallets as shown in Fig. 1, are responsible for managing the material flow within warehouses. To optimize the flow, tracking systems have been developed to enable the traceability of such carriers. These tracking systems mostly use Global Returnable Asset Identifiers (GRAI), which are physically attached to the carrier. GRAIs can be barcodes, Radio Frequency Identification Tags (RFID), or Data Matrices [1]. However, while GRAIs improve traceability, their attachment entails additional costs and is susceptible to damage. Therefore, recent research [2] aims to remove the need for GRAIs by automating pallet re- identification through the naturally occurring patterns on their wooden blocks, as shown in Fig. 2. The approach utilizes machine learning architectures that were initially designed to re-identify pedestrians for surveillance purposes [3]. In the context of Computer Vision, re-identification describes the process of correctly distinguishing an object instance from other similar objects across multiple images. Instead of tracking pallets by manually applying GRAIs onto them, every logistics distribution center can set up cameras to capture images of the pallet blocks. Every time a pallet passes by a camera, an image of its blocks will be taken using object detection, which can then be used to re-identify the pallet. While the research in [2] reached a mean Average Precision (mAP) of 98%, it still comes with some limitations, such that it does not allow conclusions about the general feasibility of automatic re-identification of pallet blocks. Firstly, the dataset used is limited in size (5,020 images) [4]. Secondly, the dataset is homogeneous, as all pallet blocks are pristine and non-branded, as seen in Fig. 2. Fig. 1.

The European (EPAL) Pallet. Highlighted in red: The pallet blocks that are necessary for the serialization and re-identification of a pallet.

Fig. 2.

Three example images of the same pallet block from [4] (ID 1).