mehmetkeremturkcan/YOLO11xxl

This is a YOLO model based on the YOLO11x architecture. It achieves an mAP50-95 of 57.3% on the COCO val 2017 dataset and achieves 30FPS on last-generation commercial GPUs, a state-of-the-art for models that can achieve real-time object detection with modern GPUs.

Importantly, this model does not use external data for pretraining unlike other SOTA-like models in the modern object detection space, and thus is uncontaminated with validation data. We are in the process of training the next generation of models with external data as a next step.

Setup

pip install ultralytics

Test

from ultralytics import YOLO
model = YOLO('yolo11xxl.pt')
results = model.predict('example.png', augment=True, imgsz=1280)

Metrics:

Detection: COCO Val

(AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.573
(AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.742
(AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.627
(AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.440
(AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.618
(AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.717
(AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.412
(AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.694
(AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.753
(AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.631
(AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.786
(AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.867
(AR) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.934
(AR) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.828

Inference Time

RTX 4080

Format	Inference time (ms/im)	FPS
PyTorch	31.57	31.68
TorchScript	24.84	40.26
ONNX	38.49	25.98
TensorRT	18.19	54.96
TensorRT (float16)	6.33	157.95
TensorRT (int8)	4.5	222.14

Acknowledgements

This work was supported in part by the National Science Foundation under grant EEC-2133516, with compute grants from Empire AI Consortium and NVIDIA Academic Grant program.