On the Effectiveness of Heterogeneous Ensemble Methods for Re-identification

TL;DR

This paper addresses industrial re-identification by replacing single, complex siamese networks with heterogeneous ensembles of simpler, faster sub-models. It introduces five or more diverse sub-models (image-based, graph-based, linear-quantile, brightness, color-based, and color-variance variants) and five ensemble transformations (Concatenation, NN-triplet stacking, Weighted Triplet, Weighted Accuracy, and Majority Vote), demonstrating that ensembles significantly surpass individual models. Across pallet-block and galvanized-metal datasets, the concatenation ensemble achieves strong Rank-1 and Rank-10 performance, approaching or exceeding state-of-the-art baselines while reducing training time, illustrating practical benefits for hardware-constrained industrial settings. The work also discusses trustworthiness and reproducibility, suggesting that heterogeneous ensembles offer robustness and interpretability advantages for re-identification tasks, with potential applicability to other contrastive learning problems.

Abstract

In this contribution, we introduce a novel ensemble method for the re-identification of industrial entities, using images of chipwood pallets and galvanized metal plates as dataset examples. Our algorithms replace commonly used, complex siamese neural networks with an ensemble of simplified, rudimentary models, providing wider applicability, especially in hardware-restricted scenarios. Each ensemble sub-model uses different types of extracted features of the given data as its input, allowing for the creation of effective ensembles in a fraction of the training duration needed for more complex state-of-the-art models. We reach state-of-the-art performance at our task, with a Rank-1 accuracy of over 77% and a Rank-10 accuracy of over 99%, and introduce five distinct feature extraction approaches, and study their combination using different ensemble methods.

Figures (10)

• Figure 1: Excerpts of the used datasets. a) pallet block dataset rutinowski_jerome_2021_6353714, b) galvanized metal dataset rutinowski_jerome_2022_7386956
• Figure 2: Rank-k accuracy for the ensemble sub-models. The highlighted areas represent the highest and lowest scores per individual fold. a) pallet block dataset, b) metal dataset
• Figure 3: Rank-1, Rank-6, and Rank-10 accuracy for the five different ensemble approaches compared to our most effective non-ensemble model (the graph-based model). a) pallet block dataset, b) metal dataset
• Figure 4: Relative uncertainty (accuracy uncertainty divided by accuracy) of the Rank-10 accuracy for different models.
• Figure 5: Rank-1 to Rank-4 gallery matches (b) to a query image (a), for three sub-models. The Image ID is indicated in the corner, correct matches are framed in green, incorrect ones in red.