Skip to content

基于CenterNet训练的目标检测&人脸对齐&姿态估计模型

Notifications You must be signed in to change notification settings

bleakie/CenterMulti

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

23 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CenterMulti代码实现参考以下

Object detection, 3D detection, and pose estimation using center point detection:

Objects as Points,
Xingyi Zhou, Dequan Wang, Philipp Krähenbühl,

2021.03.16:centernet2横空出世

COCO

Model COCO val mAP FPS
CenterNet-S4_DLA_8x 42.5 71
CenterNet2_R50_1x 42.9 24
CenterNet2_X101-DCN_2x 49.9 8
CenterNet2_R2-101-DCN-BiFPN_4x+4x_1560_ST 56.1 5
CenterNet2_DLA-BiFPN-P5_24x_ST 49.2 38

2020.04.09:基于centernet的the-state-of-the-art目标跟踪方法

tracker

2020.03.25: > 更强大的centernet优化版本,resnet50+without DCN+mAP=35.7(3.1% )↑

2019.06.10: > CenterNet code

1. Backebone

  • Strong: 增加支持mobilenetV2,mobilenetV3,efficientdet,shufflenetv2,部分网络需要支持DCNv2.

performance

  • Data process: 添加widerface转coco格式,参见root/data_process.

2. Purpose

2.1 Object Detection

  • 类别: 可支持行人、人脸、车辆、缺陷等检测,只需要修改数据加载即可
Backbone AP / FPS Flip AP / FPS Multi-scale AP / FPS
Hourglass-104 40.3 / 14 42.2 / 7.8 45.1 / 1.4
DLA-34 37.4 / 52 39.2 / 28 41.7 / 4
ResNet-101 34.6 / 45 36.2 / 25 39.3 / 4
ResNet-18 28.1 / 142 30.0 / 71 33.2 / 12

All models and details are available in > CenterNet MODEL_ZOO

shoulder defect

2.2 keypoint or pose

  • 姿态估计or关键点检测: 修改keypoint的数量及coco加载keypoint的格式可针对性训练多种形式的pose(如landmark等)

姿态估计参考centerpose

Backbone AP FPS TensorRT Speed Download
DLA-34 62.7 23 - model
Resnet-50 54.5 28 33 model
MobilenetV3 46.0 30 50 model
ShuffleNetV2 43.9 25 - model
High Resolution 57.1 16 - model
HardNet 45.6 30 - model
Darknet53 34.2 30 - model

face_landmark

2.3 model

centerface/shoulder/defect模型 提取码: u3pq

  • defect: defect模型基于mobilenetv2训练,由于部分数据标定不准,所以结果会有偏差,建议只供pre-train.

  • centerface: 该版本的centerface是基于修改的centernet训练,训练数据参照widerface,其中对质量不好的face做了过滤,使其更适合人脸识别的工程应用,模型有两个,分别是3.5M和8.9M.

centerface的训练:例如修改lib/datasets/coco_hp.py里num_joints = 5;flip_idx = [[0, 1], [3, 4]]以及整个项目里17的关节点数全部置换成5,dets[39:51]这类全部换成dets[15:20]等

3. TensorRT

  1. torch转onnx
python convert2onnx.py
  1. onnx转TensorRT
python demo_tensorrt.py
  1. 检测框架支持的TensorRT

TensorRT C++

    #shoulder检测模型支持该框架加速(不需要DCNs),total runtime = 3.82147 ms
    #在include/ctdetConfig.h里添加以下,然后cmake即可
    constexpr static int input_w = 512 ;
    constexpr static int input_h = 512 ;
    constexpr static int channel = 3 ;
    constexpr static int classNum = 1 ;
    constexpr static float mean[]= {0.408, 0.447, 0.470};
    constexpr static float std[] = {0.289, 0.274, 0.278};
    constexpr static char *className[]= {(char*)"shoulder"};

Citation

If you find this project useful for your research, please use the following BibTeX entry.

@contact{[email protected],
  title={Objects as Points},
  author={bleakie},
  year={2019}
}

About

基于CenterNet训练的目标检测&人脸对齐&姿态估计模型

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages