PennSyn2Real: Training Object Recognition Models Without Human Labeling

Scalable training data generation is a critical problem in deep learning. We propose PennSyn2Real - photo-realistic synthetic dataset consisting of more than 100 000 4K images 20 types micro aerial vehicles (MAVs). The can be used to generate arbitrary numbers for high-level computer vision tasks such as MAV detection and classification. Our framework bootstraps chroma-keying, mature cinematography technique, with motion tracking system providing artifact-free curated annotated images. system, therefore, allows object orientations lighting controlled. This easy set up applied broad range objects, reducing the gap between real-world data. show that generated using this directly train CNN models common recognition segmentation. demonstrate competitive performance comparison only real Furthermore, bootstrapping few-shot learning significantly improve overall performance, number required samples achieve desired accuracy.