Exploring the Use of Synthetic Data to Improve Pedestrian Detection in CNN

• Rubens Montanha PUCRS
• Soraia Musse PUCRS

Resumo

In this paper, we investigate the use of synthetic data generation to enhance the accuracy of a pedestrian detection task utilizing a popular Deep Convolutional Neural Network framework (YOLOv10). Using virtual characters, we develop a model for creating computer-generated images, simulating these characters in a virtual environment, and automatically annotating the data based on their position. The synthetic data is then mixed with the real data to fine-tune YOLO and test its capacity to detect pedestrians. We experiment using MOT20 datasets, a benchmark for pedestrian detection, and create a combination of mixing real and CG data. Results show that fine-tuning with Computer graphics data can positively impact the neural network's Precision, Recall, and F1-score metrics, with promising potential in improving model performance and robustness, especially in scenarios with limited real data.

Palavras-chave: YOLO, Measurement, Pedestrians, Computational modeling, Virtual environments, Computer graphics, Data models, Robustness, Convolutional neural networks, Synthetic data