A synthetic data approach for object detection in super low-resolution images

What is it about?

This paper presents a synthetic data approach to train object detection models to address the challenges with object detection in super low-resolution images. With a particular emphasis on person detection, the study uses 28 photorealistic 3D models of individuals, optimised for efficient rendering and minimal memory consumption. These models are seamlessly integrated into a 3D terrain model, mimicking diverse real-world situations. To ensure scalability and diversity, the methodology incorporates domain randomisation techniques, encompassing variations in factors like lighting conditions, seasonal effects, camera angles, lens specifications, and different image resolutions. The process of dataset generation is automated through a Python script in Blender, offering systematic scene configuration and camera positioning. The dataset created consists of 10,560 images across four resolutions. The evaluation was carried out using popular object detection algorithms, including Faster RCNN and RetinaNet, within the Detectron2 framework. Results highlight the effectiveness of synthetic datasets in training and testing object detection algorithms, showcasing visual comparisons, Average Precision (AP) metrics, and training performance statistics. Notably, RetinaNet outperforms Faster RCNN, achieving higher accuracy.

Featured Image

Why is it important?

This research offers invaluable insights into synthetic dataset generation and its application for object detection in low-resolution images.