We developed “YOLO Phantom” for detection in low-light conditions and occluded scenarios within resource-constrained IoT applications. We proposed the novel "Phantom Convolution," which enables YOLO Phantom to achieve comparable accuracy to YOLOv8n with a 43% reduction in parameters and size, resulting in a 19% reduction in GFLOPs. By employing transfer learning on our multimodal dataset, the model demonstrates enhanced vision capabilities in adverse conditions. Our Raspberry Pi IoT platform equipped with noIR cameras and integration with AWS IoT Core and SNS showcases a substantial 17% and 14% boost in frames per second for thermal and RGB data detection, respectively, compared to the baseline YOLOv8n model.
To know more about the MODIPHY please refer to the preprint available in
arXiv@article{mukherjee2024modiphy,
title={MODIPHY: Multimodal Obscured Detection for IoT using PHantom Convolution-Enabled Faster YOLO},
author={Mukherjee, Shubhabrata and Beard, Cory and Li, Zhu},
journal={arXiv preprint arXiv:2402.07894},
year={2024}
}
