Fusion or Confusion? a Look at Dataset Pooling for Infrared Object Detection

Abstract

Data pooling, by fusing individual datasets, aims to improve the generalization and robustness of object detectors. While this approach offers clear benefits, it also introduces challenges and may sometimes lead to counterproductive results. Assessing its effectiveness is challenging due to the difficulty in quantifying dataset informativeness. A common yet cumbersome method is to benchmark detector performance while optimizing the data pool composition. In this paper, we conduct a comprehensive evaluation of data pooling for object detection using infrared datasets, focusing on 'vehicles' as a reference class. While challenges exist across different spectral domains, infrared imagery presents unique complexities due to its reliance on pre-processing, dataset heterogeneity, and image quality variations. Since pre-processing addresses issues such as temperature variability, sensor noise, and dataset inconsistencies, we further examine its impact on data pooling. Additionally, we evaluate zero-shot performance on single dataset models. The paper provides a structured assessment of data pooling effectiveness through extensive experiments on seven publicly available datasets, offering insights into its practical implications.