The Roboflow ML team has been actively working on RF-DETR, a real-time, transformer-based object detection model architecture. The model architecture is now public and open source (Apache 2.0). RF-DETR-large is the first real-time model to exceed 60 AP on the Microsoft COCO benchmark [1] and achieve strong speeds compared to other models at the base size. On an NVIDIA T4 GPU, RF-DETR-Base achieves 160 FPS. RF-DETR is designed to transfer well to identify real-world objects that aren’t usually found in common training datasets, such as those found in industrial environments, wildlife settings, lab images, thermal, and novel research areas. This is measured using the new RF100-VL benchmark [2] developed in partnership with researchers from CMU. If you try out the model, let us know! We have a fine-tuning guide to get you started with training models. [3] [2]: https://github.com/roboflow/rf100-vl [3]: https://blog.roboflow.com/train-rf-detr-on-a-custom-dataset/ |