A Study on the Impact of the Occlusion Level Distribution in Datasets on the Robustness of Target Detection

Detection of occluded targets remains a challenge in computer vision. Collecting a large amount of occlusion datasets is costly, and current data augmentation methods for simulating occlusion do not consider the impact of occlusion level distribution. Various schemes were designed based on the diversity and the mean value of occlusion level to study the impact of occlusion level distribution in datasets on the robustness of target detection. To facilitate experiments, a novel benchmarking dataset NPOD-L (a Novel PASCAL3D+ Occlusion Dataset for Learning) with occlusion level labels is introduced to simulate occlusion with different levels. The new dataset contains samples of 10,200 images of 12 categories. By training on each scheme and subsequently testing on a test set using the YOLOv8 model, we investigated the effects of those schemes on thse model's performance. The results show that there exists an optimized proportioning scheme for the occlusion degree distribution such that the mAP(mean Average Precision) of the dataset is reached in most cases. We validate the conclusions' generality through experiments on NOO-DLES1 and TOOL-D occlusion datasets, showcasing case-specific distributions' benefits and practical efficacy. This serves as a useful reference for enhancing model performance via data augmentation and regularization techniques.