/yolov3mobile

Mobile YOLOv3 object detector(person and face)

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

yolov3mobile

This project originated from the excellent ultralytics/yolov3 repository and aims to implement light mobile object detector.

Imgur Credit: Ayoosh Kathuria

Data Preparation

  • COCO
  • COCO_person, 训练集60000张, 验证集4115张
  • WiderPerson, 训练集8000张, 验证集1000张
  • Citypersons, 训练集2112张, 验证集487张
  • ICCVWider, 包括监控(sur)和自动驾驶(ad), 训练集58026张, 验证集3233张
  • WIDERFace, 训练集12881张, 验证集3220张
  • AI-Challenger single, 训练集22446张, 验证集1500张
  • AI-Challenger, 训练集210000张, 验证集30000张

Baidu Drive(code:9mjn)

Type Dataset Train data Val data
CC COCO_person 60000 4115
WP WiderPerson 8000 1000
CP City_person 2112 487
Wsur ICCV_sur 5563 2369
Wad ICCV_ad 52436 864
ACS AI-Challenger 22446 1500
AC AI-Challenger 210000 30000
WF WIDERFace 12880 3226

Benchmark

COCO detection

Network Input size FLOPS Params COCO mAP@0.5 Weight size
MobileNetV2-YOLOv3-Lite 320 1.86G 1.97M 40.1 7.8MB
MobileNetV2-YOLOv3-Nano 320 0.55G 0.87M 30.4 3.6MB

Person detection(single class)

Network Input size Train data Val data FLOPS Params mAP@0.5 Weight size
MobileNetV2-YOLOv3-Nano 320 CC+WP+CP CC+WP+CP 0.48G 0.71M 50.3 2.9M
MobileNetV2-YOLOv3-Nano-3yolo 320 CC+WP+CP CC+WP+CP 0.69G 0.76M 56.3 3.2M
YOLOv3-Ultra 320 CC+WP+CP CC+WP+CP 0.11G 0.1M 46.1 515K
YOLOv3-Ultra 320 CC+WP+CP+Wsur CC+WP+CP 0.11G 0.1M 45.4 515K
YOLOv3-Ultra 320 CC CC 0.11G 0.1M 48.1 515K
MobileNetV2-YOLOv3-Nano 320 AC AC 0.48G 0.71M 57.0 2.9M
YOLOv3-Ultra 320 AC AC 0.11G 0.1M 44.5 515K

Face detection(single class)

Network Input size Train data Val data FLOPS Params Weight size mAP@0.5 Weight size Easy Medium Hard
Retinaface-m0.25 320 WF WF - - 1.68M - 1.68M 74.5 55.3 23.2
YOLOv3-Ultra 320 WF WF 110M 0.1M 515K 35.8 516K 75.507 71.034 43.083
MobileNetV2-YOLOv3-Nano-3yolo 320 WF WF 0.69G 0.76M 3.2M 39.2 3.2M 83.464 79.303 47.205