YOLO-ReT: Towards High Accuracy Real-time Object Detection on Edge GPUs

As object detection models have improved in performance and grown in size, mapping them to edge devices such as Jetson involves tradeoffs between model size and accuracy. Despite these reductions in model size, the results do not necessarily improve speeds or produce optimal designs. YOLO-ReT: Towards High Accuracy Real-time Object Detection on Edge GPUs proposes a novel GPU-friendly transfer learning module improving on accuracy and execution speeds by employing Backbone Truncation and Multi-Scale Feature Interaction. The measured performance improvements of these modules were tested using the COCO and Pascal VOC datasets on Jetson Nano, Jetson Xavier NX and Jetson AGX Xavier. Using various backbones and input resolutions, the team behind this project demonstrate models from the YOLO-ReT family with lower latency and better accuracy.