Skip to content

Latest commit

 

History

History
366 lines (317 loc) · 29.9 KB

README.md

File metadata and controls

366 lines (317 loc) · 29.9 KB
Raw

简体中文 | English

人脸检测SCRFD

目录

1. 简介

SCRFD(Sample and Computation Redistribution for Efficient Face Detection)是一种基于FCOS 的人脸检测算法，该算法在2021年5月推出。它被设计为一个高效和高精度的人脸检测器，其速度和准确性相较于其他现有算法都有显著提高。

论文地址 (https://arxiv.org/pdf/2105.04714.pdf)

源码地址 (https://github.com/deepinsight/insightface/tree/master/detection/scrfd)

2. 特性

  • 支持BM1688/CV186X(SoC)、BM1684X(x86 PCIe、SoC)、BM1684(x86 PCIe、SoC)
  • 支持FP32、FP16(BM1684X/BM1688/CV186X)、INT8模型编译和推理
  • 支持基于BMCV预处理的C++推理
  • 支持基于OpenCV和BMCV预处理的Python推理
  • 支持单batch模型推理
  • 支持图片和视频测试

3. 准备模型与数据

建议使用TPU-MLIR编译BModel，Pytorch模型在编译前要导出成onnx模型，如果您使用的tpu-mlir版本>=v1.3.0（即官网v23.07.01），可以直接使用torchscript模型。具体可参考SCRFD模型导出方法。

​同时，您需要准备用于测试的数据集，如果量化模型，还要准备用于量化的数据集。 本例程提供了一种性能和精度上较高的模型， scrfd_10g_kps.onnx 。您都可以使用MLIR工具链转出为对应的bmodel模型。

如果您想使用其他模型，您可以访问 源码地址 进行下载，并参考 SCRFD模型导出方法 进行导出。

同时，您需要准备用于测试的数据集，如果量化模型，还要准备用于量化的数据集。

本例程在scripts目录下提供了相关模型和数据集的下载脚本download.sh，您也可以自己准备模型和数据集，并参考4. 模型编译进行模型转换。

# 安装unzip，若已安装请跳过，非ubuntu系统视情况使用yum或其他方式安装
sudo apt install unzip
chmod -R +x scripts/
./scripts/download.sh

下载的模型包括

./models
.
├── BM1684                                   # 使用TPU-MLIR编译，用于BM1684的 BModel
│   ├── scrfd_10g_kps_fp32_1b.bmodel
│   ├── scrfd_10g_kps_int8_1b.bmodel
│   ├── scrfd_10g_kps_int8_4b.bmodel
├── BM1684X                                  # 使用TPU-MLIR编译，用于BM1684X的 BModel
│   ├── scrfd_10g_kps_fp16_1b.bmodel
│   ├── scrfd_10g_kps_fp32_1b.bmodel
│   ├── scrfd_10g_kps_int8_1b.bmodel
│   ├── scrfd_10g_kps_int8_4b.bmodel
├── BM1688                                   # 使用TPU-MLIR编译，用于BM1688的 BModel
│   ├── scrfd_10g_kps_fp16_1b_2core.bmodel
│   ├── scrfd_10g_kps_fp16_1b.bmodel
│   ├── scrfd_10g_kps_fp32_1b_2core.bmodel
│   ├── scrfd_10g_kps_fp32_1b.bmodel
│   ├── scrfd_10g_kps_int8_1b_2core.bmodel
│   ├── scrfd_10g_kps_int8_1b.bmodel
│   ├── scrfd_10g_kps_int8_4b_2core.bmodel
│   ├── scrfd_10g_kps_int8_4b.bmodel
├── CV186X                                   # 使用TPU-MLIR编译，用于CV186X的 BModel
│   ├── scrfd_10g_kps_fp16_1b.bmodel
│   ├── scrfd_10g_kps_fp32_1b.bmodel
│   ├── scrfd_10g_kps_int8_1b.bmodel
│   ├── scrfd_10g_kps_int8_4b.bmodel
└── onnx                                     # 导出的onnx模型
    ├── scrfd_10g_kps_1b.onnx
    ├── scrfd_10g_kps_4b.onnx

下载的数据包括：

./datasets
├── face_det.mp4                     # 测试视频
├── test                             # 测试图片
│   ├── men.jpg
│   └── selfie.jpg
└── WIDER_val                        # 精度评估数据集
    └── images

4. 模型编译

模型编译前需要安装TPU-MLIR，具体可参考TPU-MLIR环境搭建。安装好后需在TPU-MLIR环境中进入例程目录。使用TPU-MLIR将onnx模型编译为BModel，具体方法可参考《TPU-MLIR快速入门手册》的“3. 编译ONNX模型”(请从算能官网相应版本的SDK中获取)。

  • 生成FP32 BModel

​本例程在scripts目录下提供了TPU-MLIR编译FP32 BModel的脚本，请注意修改gen_fp32bmodel_mlir.sh中的onnx模型路径、生成模型目录和输入大小shapes等参数，并在执行时指定BModel运行的目标平台（支持BM1684X/BM1688/CV186X），如：

./scripts/gen_fp32bmodel_mlir.sh bm1684 #bm1684x/bm1688/cv186x

​执行上述命令会在models/BM1684等文件夹下生成scrfd_10g_kps_fp32_1b.bmodel文件，即转换好的FP32 BModel。

  • 生成FP16 BModel

​本例程在scripts目录下提供了TPU-MLIR编译FP16 BModel的脚本，请注意修改gen_fp16bmodel_mlir.sh中的onnx模型路径、生成模型目录和输入大小shapes等参数，并在执行时指定BModel运行的目标平台（支持BM1684X/BM1688/CV186X），如：

./scripts/gen_fp16bmodel_mlir.sh bm1684x #bm1688/cv186x

​执行上述命令会在models/BM1684X/等文件夹下生成scrfd_10g_kps_fp16_1b.bmodel文件，即转换好的FP16 BModel。

  • 生成INT8 BModel

​本例程在scripts目录下提供了量化INT8 BModel的脚本，请注意修改gen_int8bmodel_mlir.sh中的onnx模型路径、生成模型目录和输入大小shapes等参数，在执行时输入BModel的目标平台（支持BM1684X/BM1688/CV186X），如：

./scripts/gen_int8bmodel_mlir.sh bm1684 #bm1684x/bm1688/cv186x

​上述脚本会在models/BM1684等文件夹下生成scrfd_10g_kps_int8_1b.bmodel等文件，即转换好的INT8 BModel。

5. 例程测试

6. 精度测试

6.1 测试方法

首先，参考C++例程或Python例程推理要测试的数据集，生成预测的txt文件夹，注意修改数据集(datasets/WIDER_val)和相关参数(conf_thresh=0.02、nms_thresh=0.45以及--eval=True)。
然后，使用tools目录下的evaluation.py脚本，将测试生成的txt文件夹与测试集标签ground_truth文件夹进行对比，计算出人脸检测的评价指标，命令如下：

cd tools
pip3 install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
python3 setup.py build_ext --inplace
python3 evaluation.py --pred ./prediction_dir --gt ground_truth

具体测试方法，请参考精度测试

6.2 测试结果

在WIDER FACE数据集上，官方SCRFD_10G_KPS模型的精度测试结果是：Easy: 0.9540, Medium: 0.9401, Hard: 0.8280，本例程的精度测试结果如下表所示：

测试平台 测试程序 测试模型  Easy Medium  Hard
SE5-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 0.940 0.924 0.800
SE5-16 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 0.913 0.904 0.783
SE5-16 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 0.929 0.916 0.796
SE5-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 0.939 0.921 0.784
SE5-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 0.912 0.900 0.767
SE5-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 0.926 0.912 0.778
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 0.936 0.917 0.764
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 0.835 0.825 0.659
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 0.864 0.847 0.677
SE5-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 0.936 0.917 0.764
SE5-16 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 0.835 0.825 0.659
SE5-16 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 0.864 0.847 0.677
SE7-32 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 0.940 0.924 0.800
SE7-32 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 0.940 0.924 0.800
SE7-32 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 0.939 0.923 0.796
SE7-32 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 0.939 0.923 0.799
SE7-32 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 0.939 0.921 0.786
SE7-32 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 0.939 0.921 0.786
SE7-32 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 0.938 0.919 0.783
SE7-32 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 0.937 0.919 0.783
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 0.937 0.917 0.772
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 0.937 0.917 0.772
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 0.885 0.863 0.689
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 0.887 0.865 0.691
SE7-32 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 0.937 0.917 0.772
SE7-32 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 0.937 0.917 0.772
SE7-32 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 0.885 0.863 0.689
SE7-32 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 0.887 0.864 0.690
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 0.940 0.924 0.800
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 0.940 0.924 0.800
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 0.938 0.923 0.798
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 0.939 0.923 0.798
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 0.938 0.919 0.780
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 0.938 0.919 0.780
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 0.936 0.917 0.776
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 0.936 0.917 0.776
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 0.936 0.916 0.766
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 0.936 0.916 0.766
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 0.886 0.864 0.687
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 0.886 0.864 0.687
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 0.930 0.912 0.764
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 0.935 0.915 0.765
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 0.885 0.863 0.687
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 0.885 0.863 0.686
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b_2core.bmodel 0.940 0.924 0.800
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp16_1b_2core.bmodel 0.940 0.924 0.800
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_1b_2core.bmodel 0.938 0.923 0.798
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_4b_2core.bmodel 0.939 0.923 0.798
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b_2core.bmodel 0.938 0.919 0.780
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp16_1b_2core.bmodel 0.938 0.919 0.780
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b_2core.bmodel 0.936 0.917 0.776
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b_2core.bmodel 0.936 0.917 0.776
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b_2core.bmodel 0.936 0.916 0.766
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b_2core.bmodel 0.936 0.916 0.766
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b_2core.bmodel 0.886 0.864 0.687
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b_2core.bmodel 0.886 0.864 0.687
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b_2core.bmodel 0.931 0.913 0.764
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp16_1b_2core.bmodel 0.931 0.912 0.764
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_1b_2core.bmodel 0.884 0.862 0.686
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_4b_2core.bmodel 0.885 0.863 0.686
SE9-8 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 0.940 0.924 0.800
SE9-8 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 0.940 0.924 0.800
SE9-8 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 0.938 0.923 0.798
SE9-8 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 0.939 0.923 0.799
SE9-8 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 0.938 0.919 0.778
SE9-8 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 0.938 0.919 0.778
SE9-8 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 0.936 0.917 0.776
SE9-8 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 0.936 0.917 0.776
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 0.936 0.916 0.765
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 0.936 0.916 0.765
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 0.885 0.863 0.687
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 0.886 0.864 0.688
SE9-8 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 0.936 0.916 0.765
SE9-8 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 0.936 0.916 0.765
SE9-8 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 0.885 0.863 0.687
SE9-8 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 0.885 0.863 0.687

7. 性能测试

7.1 bmrt_test

使用bmrt_test测试模型的理论性能：

# 请根据实际情况修改要测试的bmodel路径和devid参数
bmrt_test --bmodel models/BM1684X/scrfd_10g_kps_fp32_1b.bmodel

测试结果中的calculate time就是模型推理的时间，多batch size模型应当除以相应的batch size才是每张图片的理论推理时间。 测试各个模型的理论推理时间，结果如下：

测试模型 calculate time(ms)
BM1684/scrfd_10g_kps_fp32_1b.bmodel 20.082
BM1684/scrfd_10g_kps_int8_1b.bmodel 16.265
BM1684/scrfd_10g_kps_int8_4b.bmodel 5.024
BM1684X/scrfd_10g_kps_fp16_1b.bmodel 3.791
BM1684X/scrfd_10g_kps_fp32_1b.bmodel 34.830
BM1684X/scrfd_10g_kps_int8_1b.bmodel 2.645
BM1684X/scrfd_10g_kps_int8_4b.bmodel 2.537
BM1688/scrfd_10g_kps_fp16_1b.bmodel 45.524
BM1688/scrfd_10g_kps_fp16_1b_2core.bmodel 31.586
BM1688/scrfd_10g_kps_fp32_1b.bmodel 323.095
BM1688/scrfd_10g_kps_fp32_1b_2core.bmodel 190.639
BM1688/scrfd_10g_kps_int8_1b.bmodel 13.398
BM1688/scrfd_10g_kps_int8_1b_2core.bmodel 11.044
BM1688/scrfd_10g_kps_int8_4b.bmodel 12.720
BM1688/scrfd_10g_kps_int8_4b_2core.bmodel 7.188
CV186X/scrfd_10g_kps_fp16_1b.bmodel 42.652
CV186X/scrfd_10g_kps_fp32_1b.bmodel 317.354
CV186X/scrfd_10g_kps_int8_1b.bmodel 13.034
CV186X/scrfd_10g_kps_int8_4b.bmodel 12.323

测试说明：

  1. 性能测试结果具有一定的波动性；
  2. calculate time已折算为平均每张图片的推理时间；
  3. SoC和PCIe的测试结果基本一致。

7.2 程序运行性能

参考C++例程或Python例程运行程序，并查看统计的解码时间、预处理时间、推理时间、后处理时间。

在不同的测试平台上，使用不同的例程、模型测试datasets/WIDER_val，conf_thresh=0.5，nms_thresh=0.5，性能测试结果如下：

测试平台 测试程序 测试模型 decode_time preprocess_time inference_time postprocess_time
SE5-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 13.56 24.75 25.05 8.39
SE5-16 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 11.01 25.92 21.25 8.35
SE5-16 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 11.19 26.34 8.68 8.27
SE5-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 3.70 3.84 22.14 8.75
SE5-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 3.68 3.83 18.31 8.51
SE5-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 3.45 3.63 6.08 8.40
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 4.42 0.97 20.06 8.55
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 4.40 0.97 16.25 8.18
SE5-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 4.20 0.91 5.03 8.54
SE5-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 3.23 3.93 20.48 8.29
SE5-16 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 3.21 3.93 16.66 8.22
SE5-16 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 3.06 3.72 5.28 7.79
SE7-32 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 11.01 25.01 40.31 8.65
SE7-32 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 11.21 25.73 9.24 8.68
SE7-32 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 11.42 25.77 8.07 8.55
SE7-32 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 11.09 27.01 6.66 8.38
SE7-32 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 4.82 2.93 36.86 8.66
SE7-32 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 3.02 2.93 5.85 8.63
SE7-32 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 3.02 2.91 4.69 8.66
SE7-32 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 2.84 2.75 3.62 8.49
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 3.94 0.87 34.85 8.46
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 3.94 0.87 3.81 8.46
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 3.92 0.87 2.65 8.42
SE7-32 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 3.70 0.84 2.53 8.54
SE7-32 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 2.73 3.15 35.29 8.48
SE7-32 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 2.73 3.17 4.25 8.56
SE7-32 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 2.72 3.17 3.11 8.53
SE7-32 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 2.54 3.09 2.79 7.82
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 45.48 33.19 170.02 11.41
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 49.88 32.64 30.87 11.49
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 50.62 32.52 14.32 11.57
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 45.14 35.53 12.01 11.23
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 9.31 6.89 165.62 11.49
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 12.83 6.89 26.40 11.55
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 12.73 6.88 9.98 11.55
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 9.59 6.56 8.09 11.13
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 8.38 2.43 162.74 11.78
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 8.77 2.43 23.60 11.77
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 6.41 2.43 7.19 11.58
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 6.03 2.35 6.76 11.93
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 11.47 6.84 163.27 11.98
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 11.45 6.84 24.09 11.93
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 10.16 6.83 7.67 12.02
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 5.65 6.66 6.96 10.95
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp32_1b_2core.bmodel 43.36 33.20 103.05 11.41
SE9-16 scrfd_opencv.py scrfd_10g_kps_fp16_1b_2core.bmodel 48.23 32.99 23.20 11.48
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_1b_2core.bmodel 50.92 33.15 13.04 11.50
SE9-16 scrfd_opencv.py scrfd_10g_kps_int8_4b_2core.bmodel 45.26 36.02 9.06 11.26
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp32_1b_2core.bmodel 11.12 6.91 98.64 11.47
SE9-16 scrfd_bmcv.py scrfd_10g_kps_fp16_1b_2core.bmodel 5.53 6.89 19.00 11.53
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_1b_2core.bmodel 5.43 6.88 8.74 11.53
SE9-16 scrfd_bmcv.py scrfd_10g_kps_int8_4b_2core.bmodel 5.19 6.58 5.22 11.16
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b_2core.bmodel 7.16 2.43 95.78 11.67
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b_2core.bmodel 6.78 2.43 16.20 11.72
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_1b_2core.bmodel 6.21 2.43 5.92 11.72
SE9-16 scrfd_bmcv.soc scrfd_10g_kps_int8_4b_2core.bmodel 5.99 2.35 3.89 11.92
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp32_1b_2core.bmodel 5.81 6.83 96.29 12.05
SE9-16 scrfd_sail.soc scrfd_10g_kps_fp16_1b_2core.bmodel 5.37 6.82 16.69 12.02
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_1b_2core.bmodel 4.83 6.82 6.41 11.95
SE9-16 scrfd_sail.soc scrfd_10g_kps_int8_4b_2core.bmodel 4.65 6.66 4.09 10.78
SE9-8 scrfd_opencv.py scrfd_10g_kps_fp32_1b.bmodel 45.91 34.23 324.44 11.54
SE9-8 scrfd_opencv.py scrfd_10g_kps_fp16_1b.bmodel 71.27 33.46 49.85 11.56
SE9-8 scrfd_opencv.py scrfd_10g_kps_int8_1b.bmodel 78.83 32.81 20.13 11.57
SE9-8 scrfd_opencv.py scrfd_10g_kps_int8_4b.bmodel 44.62 35.96 17.73 11.44
SE9-8 scrfd_bmcv.py scrfd_10g_kps_fp32_1b.bmodel 13.58 7.24 320.40 11.52
SE9-8 scrfd_bmcv.py scrfd_10g_kps_fp16_1b.bmodel 10.34 7.25 45.68 11.69
SE9-8 scrfd_bmcv.py scrfd_10g_kps_int8_1b.bmodel 9.94 7.26 16.01 11.62
SE9-8 scrfd_bmcv.py scrfd_10g_kps_int8_4b.bmodel 15.90 6.96 13.84 11.17
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_fp32_1b.bmodel 9.60 2.57 317.28 11.74
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_fp16_1b.bmodel 10.47 2.58 42.64 11.71
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_int8_1b.bmodel 9.99 2.57 13.00 11.63
SE9-8 scrfd_bmcv.soc scrfd_10g_kps_int8_4b.bmodel 8.39 2.49 12.34 12.01
SE9-8 scrfd_sail.soc scrfd_10g_kps_fp32_1b.bmodel 8.49 7.06 317.85 12.10
SE9-8 scrfd_sail.soc scrfd_10g_kps_fp16_1b.bmodel 9.08 7.06 43.18 12.02
SE9-8 scrfd_sail.soc scrfd_10g_kps_int8_1b.bmodel 10.51 7.06 13.52 12.02
SE9-8 scrfd_sail.soc scrfd_10g_kps_int8_4b.bmodel 8.79 6.86 12.55 10.79

测试说明：

  1. 时间单位均为毫秒(ms)，统计的时间均为平均每张图片处理的时间；
  2. 性能测试结果具有一定的波动性，建议多次测试取平均值；
  3. SE5-16/SE7-32的主控处理器均为8核[email protected]，SE9-16为8核[email protected]，SE9-8为6核[email protected]，PCIe上的性能由于处理器的不同可能存在较大差异；
  4. 图片分辨率对解码时间影响较大，推理结果对后处理时间影响较大，不同的测试图片可能存在较大差异，不同的阈值对后处理时间影响较大。

8. FAQ

常见问题解答