Artificial Data Synthesis for Object Detection Training (2021)

To enhance our human detection algorithm, I devised several methods to augment or generate new data for our dataset. I began by using a cutting-edge segmentation CNN to extract human figures from the dataset images. We identified 10 distinct scenes where people were detected. For each scene, I specified the plausible areas where individuals could be positioned and developed a program to generate new scenes featuring people in different poses. To seamlessly integrate the extracted human images into these scenes, I adapted their HSV values using histogram equalization, ensuring they matched the lighting and color conditions of the scene.