yolo11face - ADD YOLOv8-pose and YOLO11-pose

This warehouse has attempted to train two model architectures in total. The first one is to train and validate the WIDERFACE dataset using only the yolov5/yolov8/yolo11 detection model architecture.

| | ARCH | GFLOPs | Easy | Medium | Hard | |:---------------------:|:------------:|:------:|:-----:|:------:|:-----:| | zjykzj/YOLO11Face | yolov5nu | 7.1 | 93.86 | 91.70 | 80.37 | | zjykzj/YOLO11Face | yolov5su | 23.8 | 95.13 | 93.47 | 84.33 | | zjykzj/YOLO11Face | yolov8s | 28.4 | 95.77 | 94.18 | 84.54 | | zjykzj/YOLO11Face | yolo11s | 21.3 | 95.55 | 93.91 | 84.85 |

The second method uses Ultralytics' pose model for joint training of faces and keypoints, and finally evaluates only the facial performance of the validation set in the original way.

Note that the facial keypoint annotation here comes from RetinaFace, which only annotated facial keypoints on the original training set. Therefore, when training the pose model, the training part of the original WIDERFACE train dataset is divided into training/validation datasets in an 8:2 ratio, and the val dataset is evaluated after training is completed.

| | ARCH | GFLOPs | Easy | Medium | Hard | |:---------------------:|:------------:|:------:|:-----:|:------:|:-----:| | zjykzj/YOLO5Face | yolov5n-v7.0 | 4.2 | 93.25 | 91.11 | 80.33 | | zjykzj/YOLO5Face | yolov5s-v7.0 | 15.8 | 94.84 | 93.28 | 84.67 | | | | | | | | | zjykzj/YOLO11Face | yolov8n-pose | 8.3 | 94.61 | 92.46 | 80.98 | | zjykzj/YOLO11Face | yolov8s-pose | 29.4 | 95.50 | 93.95 | 84.65 | | | | | | | | | zjykzj/YOLO11Face | yolo11n-pose | 6.6 | 94.62 | 92.56 | 81.02 | | zjykzj/YOLO11Face | yolo11s-pose | 22.3 | 95.72 | 94.19 | 85.24 |

During the eval phase, using VGA resolution input images (the longer edge of the input image is scaled to 640, and the shorter edge is scaled accordingly)

- Python
Published by zjykzj 12 months ago

yolo11face - Using YOLOv8-pose

Through experiments, it was found that using YOLOv8-pose can simultaneously detect faces and facial keypoints. Using the facial and keypoint datasets provided by RetinaFace, only the training dataset was used. Validation/training was split into 2:8 parts, and the facial and keypoint detectors were trained from scratch. Finally, they were evaluated on the WIDERFACE val dataset.

| | ARCH | GFLOPs | Easy | Medium | Hard | |:--------------------:|:------------:|:------:|:-----:|:------:|:-----:| | zjykzj/YOLO5Face | yolov5s-v7.0 | 15.8 | 94.84 | 93.28 | 84.67 | | zjykzj/YOLO5Face | yolov5n-v7.0 | 4.2 | 93.25 | 91.11 | 80.33 | | | | | | | | | zjykzj/YOLO8Face | yolov5su | 23.8 | 95.18 | 93.50 | 82.47 | | zjykzj/YOLO8Face | yolov5nu | 7.1 | 93.96 | 91.82 | 78.89 | | | | | | | | | zjykzj/YOLO8Face | yolov8s | 28.4 | 95.81 | 94.26 | 82.75 | | zjykzj/YOLO8Face | yolov8n | 8.1 | 94.57 | 92.55 | 78.97 | | | | | | | | | zjykzj/YOLO8Face | yolov8s-pose | 29.4 | 95.07 | 93.77 | 82.84 | | zjykzj/YOLO8Face | yolov8n-pose | 8.3 | 94.07 | 92.04 | 78.88 |

During the eval phase, using VGA resolution input images (the longer edge of the input image is scaled to 640, and the shorter edge is scaled accordingly)

- Python
Published by zjykzj about 1 year ago

yolo11face - INIT

| | ARCH | GFLOPs | Easy | Medium | Hard | |:--------------------:|:------------:|:------:|:-----:|:------:|:-----:| | zjykzj/YOLO5Face | yolov5s-v7.0 | 15.8 | 94.84 | 93.28 | 84.67 | | zjykzj/YOLO5Face | yolov5n-v7.0 | 4.2 | 93.25 | 91.11 | 80.33 | | | | | | | | | zjykzj/YOLO8Face | yolov5su | 23.8 | 95.18 | 93.50 | 82.47 | | zjykzj/YOLO8Face | yolov5nu | 7.1 | 93.96 | 91.82 | 78.89 | | | | | | | | | zjykzj/YOLO8Face | yolov8s | 28.4 | 95.81 | 94.26 | 82.75 | | zjykzj/YOLO8Face | yolov8n | 8.1 | 94.57 | 92.55 | 78.97 |

During the eval phase, using VGA resolution input images (the longer edge of the input image is scaled to 640, and the shorter edge is scaled accordingly)

During the training phase, the longer edge of the input image is scaled to 800, and the shorter edge is scaled accordingly

- Python
Published by zjykzj about 1 year ago