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

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.

• 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

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 |

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

2. onnx转TensorRT

python demo_tensorrt.py

3. 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{yangsai1991@163.com,
  title={Objects as Points},
  author={bleakie},
  year={2019}
}