Reducing Co-Occurrence Bias to Improve Classifier Explainability and Zero-Shot Detection
Hanna Witzgall, Weicheng Shen · 2022 IEEE Aerospace Conference (AERO) · 2022
This paper investigates a new component-class classification (C3) architecture to reduce the impact of co-occurrence bias that can severely degrade multi-label, component-based classification methods. Component-class classification refers to the task of classifying multiple components comprising a single object in a single image. In this paper, the objects are military vehicles with identifiable components such as turret, wheel, and gun. Multi-label architectures are a conventional deep neural network solution used for classifying images containing multiple classes or more commonly multiple class attributes. However, the addition of multiple classes inside single training images introduces the possibility of co-occurrence bias which will be shown to result in unpredictable and poor component-level classification. To mitigate this co-occurrence bias, this paper proposes to replace the multi-label classification architecture that is often used for zero-shot detection, with an architecture that is based on object detection. The result is a dramatic decrease in the effects of co-occurrence bias over that seen with the multi-label solution. Further, this paper shows that this improved component-level classification can translate into a viable approach for classifier interpretability and improve zero-shot detection applications.