Object Part Appearance Module built into Yolo for Occlusion

Despite recent advances in detection and recognition models, the development of robust and accurate algorithms for describing and classifying object visual content in real-life scenarios remains a significant challenge. This paper introduces an object part appearance module, OPA, within the framework OPA-Yolo, specifically designed for occluded object detection. Our approach combines handcrafted features with deep ones across network layers to capture the visual content parts of occulted object classes. Mapped deep features identify different regions of interest, characterizing object part appearances with a significant amount of statistical information. To enrich this feature representation with invariant local features, we train the network to learn an optimal combination of object appearance, providing it with an additional attention mechanism. The model is built on a recent YOLO object detector architecture. Evaluated on the PASCAL-VOC dataset, OPA-Yolo achieved an average detection rate of approximately 24.76% on occulted objects, with some objects having an occlusion ratio of up to 60%.