Baseline Assessment of Object Detection Models on Partially Occluded Objects
Abstract
One of the fields of computer vision commonly used in military research is object-detection. A particularly good example of this is real-time object recognition on the battlefield. Developing/evaluating these types of models requires proper object-detection and classification datasets, which are crucial for Soldiers decision-making on the battlefield. A major problem with current object-detection models is that they flounder when detecting partially occluded objects. This is because the models do not properly recognize the objects while parts of them are covered. Additionally, occlusion is not a condition that many object-detection models are designed to handle. The main objective of this work was to perform a baseline assessment of the Gonzalez-Garcia model compared with the Faster R-CNN model from Detectron2 and YOLOv5 using the PASCAL VOC 2010 dataset. Of course, this dataset contains many examples of partially occluded objects. The results from each would then be compared to determine their overall effectiveness and their accuracy on partially occluded objects. All three object-detection models seem to work well overall and somewhat well with partially occluded objects. However, none of them are very good at detecting objects in poor lighting conditions.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 14, 2022
- Accession Number
- AD1160076
Entities
People
- Darius Ii Jefferson
Organizations
- United States Army Research Laboratory