Add UltraFace Model support (#43)

* update .gitignore

* Added checking for cmake include dir

* fixed missing trt_backend option bug when init from trt

* remove un-need data layout and add pre-check for dtype

* changed RGB2BRG to BGR2RGB in ppcls model

* add model_zoo yolov6 c++/python demo

* fixed CMakeLists.txt typos

* update yolov6 cpp/README.md

* add yolox c++/pybind and model_zoo demo

* move some helpers to private

* fixed CMakeLists.txt typos

* add normalize with alpha and beta

* add version notes for yolov5/yolov6/yolox

* add copyright to yolov5.cc

* revert normalize

* fixed some bugs in yolox

* Add UltraFace Model support

.gitignore

@@ -14,3 +14,4 @@ fastdeploy/LICENSE*
 fastdeploy/ThirdPartyNotices*
 *.so*
 fastdeploy/libs/third_libs
+csrcs/fastdeploy/core/config.h

csrcs/fastdeploy/vision.h

@@ -16,6 +16,7 @@
 #include "fastdeploy/core/config.h"
 #ifdef ENABLE_VISION
 #include "fastdeploy/vision/deepcam/yolov5face.h"
+#include "fastdeploy/vision/linzaer/ultraface.h"
 #include "fastdeploy/vision/megvii/yolox.h"
 #include "fastdeploy/vision/meituan/yolov6.h"
 #include "fastdeploy/vision/ppcls/model.h"

csrcs/fastdeploy/vision/linzaer/linzaer_pybind.cc

@@ -0,0 +1,35 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/pybind/main.h"
+
+namespace fastdeploy {
+void BindLinzaer(pybind11::module& m) {
+  auto linzaer_module = m.def_submodule(
+      "linzaer",
+      "https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB");
+  pybind11::class_<vision::linzaer::UltraFace, FastDeployModel>(linzaer_module,
+                                                                "UltraFace")
+      .def(pybind11::init<std::string, std::string, RuntimeOption, Frontend>())
+      .def("predict",
+           [](vision::linzaer::UltraFace& self, pybind11::array& data,
+              float conf_threshold, float nms_iou_threshold) {
+             auto mat = PyArrayToCvMat(data);
+             vision::FaceDetectionResult res;
+             self.Predict(&mat, &res, conf_threshold, nms_iou_threshold);
+             return res;
+           })
+      .def_readwrite("size", &vision::linzaer::UltraFace::size);
+}
+}  // namespace fastdeploy

csrcs/fastdeploy/vision/linzaer/ultraface.cc

@@ -0,0 +1,220 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision/linzaer/ultraface.h"
+#include "fastdeploy/utils/perf.h"
+#include "fastdeploy/vision/utils/utils.h"
+
+namespace fastdeploy {
+
+namespace vision {
+
+namespace linzaer {
+
+UltraFace::UltraFace(const std::string& model_file,
+                     const std::string& params_file,
+                     const RuntimeOption& custom_option,
+                     const Frontend& model_format) {
+  if (model_format == Frontend::ONNX) {
+    valid_cpu_backends = {Backend::ORT};  // 指定可用的CPU后端
+    valid_gpu_backends = {Backend::ORT, Backend::TRT};  // 指定可用的GPU后端
+  } else {
+    valid_cpu_backends = {Backend::PDINFER, Backend::ORT};
+    valid_gpu_backends = {Backend::PDINFER, Backend::ORT, Backend::TRT};
+  }
+  runtime_option = custom_option;
+  runtime_option.model_format = model_format;
+  runtime_option.model_file = model_file;
+  runtime_option.params_file = params_file;
+  initialized = Initialize();
+}
+
+bool UltraFace::Initialize() {
+  // parameters for preprocess
+  size = {320, 240};
+
+  if (!InitRuntime()) {
+    FDERROR << "Failed to initialize fastdeploy backend." << std::endl;
+    return false;
+  }
+  // Check if the input shape is dynamic after Runtime already initialized,
+  is_dynamic_input_ = false;
+  auto shape = InputInfoOfRuntime(0).shape;
+  for (int i = 0; i < shape.size(); ++i) {
+    // if height or width is dynamic
+    if (i >= 2 && shape[i] <= 0) {
+      is_dynamic_input_ = true;
+      break;
+    }
+  }
+  return true;
+}
+
+bool UltraFace::Preprocess(
+    Mat* mat, FDTensor* output,
+    std::map<std::string, std::array<float, 2>>* im_info) {
+  // ultraface's preprocess steps
+  // 1. resize
+  // 2. BGR->RGB
+  // 3. HWC->CHW
+  int resize_w = size[0];
+  int resize_h = size[1];
+  if (resize_h != mat->Height() || resize_w != mat->Width()) {
+    Resize::Run(mat, resize_w, resize_h);
+  }
+
+  BGR2RGB::Run(mat);
+  // Compute `result = mat * alpha + beta` directly by channel
+  // Reference: detect_imgs_onnx.py#L73
+  std::vector<float> alpha = {1.0f / 128.0f, 1.0f / 128.0f, 1.0f / 128.0f};
+  std::vector<float> beta = {-127.0f * (1.0f / 128.0f),
+                             -127.0f * (1.0f / 128.0f),
+                             -127.0f * (1.0f / 128.0f)};  // RGB;
+  Convert::Run(mat, alpha, beta);
+
+  // Record output shape of preprocessed image
+  (*im_info)["output_shape"] = {static_cast<float>(mat->Height()),
+                                static_cast<float>(mat->Width())};
+
+  HWC2CHW::Run(mat);
+  Cast::Run(mat, "float");
+  mat->ShareWithTensor(output);
+  output->shape.insert(output->shape.begin(), 1);  // reshape to n, h, w, c
+  return true;
+}
+
+bool UltraFace::Postprocess(
+    std::vector<FDTensor>& infer_result, FaceDetectionResult* result,
+    const std::map<std::string, std::array<float, 2>>& im_info,
+    float conf_threshold, float nms_iou_threshold) {
+  // ultraface has 2 output tensors, scores & boxes
+  FDASSERT(
+      (infer_result.size() == 2),
+      "The default number of output tensor must be 2 according to ultraface.");
+  FDTensor& scores_tensor = infer_result.at(0);  // (1,4420,2)
+  FDTensor& boxes_tensor = infer_result.at(1);   // (1,4420,4)
+  FDASSERT((scores_tensor.shape[0] == 1), "Only support batch =1 now.");
+  FDASSERT((boxes_tensor.shape[0] == 1), "Only support batch =1 now.");
+
+  result->Clear();
+  // must be setup landmarks_per_face before reserve.
+  // ultraface detector does not detect landmarks by default.
+  result->landmarks_per_face = 0;
+  if (scores_tensor.dtype != FDDataType::FP32) {
+    FDERROR << "Only support post process with float32 data." << std::endl;
+    return false;
+  }
+  if (boxes_tensor.dtype != FDDataType::FP32) {
+    FDERROR << "Only support post process with float32 data." << std::endl;
+    return false;
+  }
+
+  float* scores_ptr = static_cast<float*>(scores_tensor.Data());
+  float* boxes_ptr = static_cast<float*>(boxes_tensor.Data());
+  const size_t num_bboxes = boxes_tensor.shape[1];  // e.g 4420
+  // fetch original image shape
+  auto iter_ipt = im_info.find("input_shape");
+  FDASSERT((iter_ipt != im_info.end()),
+           "Cannot find input_shape from im_info.");
+  float ipt_h = iter_ipt->second[0];
+  float ipt_w = iter_ipt->second[1];
+
+  // decode bounding boxes
+  for (size_t i = 0; i < num_bboxes; ++i) {
+    float confidence = scores_ptr[2 * i + 1];
+    // filter boxes by conf_threshold
+    if (confidence <= conf_threshold) {
+      continue;
+    }
+    float x1 = boxes_ptr[4 * i + 0] * ipt_w;
+    float y1 = boxes_ptr[4 * i + 1] * ipt_h;
+    float x2 = boxes_ptr[4 * i + 2] * ipt_w;
+    float y2 = boxes_ptr[4 * i + 3] * ipt_h;
+    result->boxes.emplace_back(std::array<float, 4>{x1, y1, x2, y2});
+    result->scores.push_back(confidence);
+  }
+
+  if (result->boxes.size() == 0) {
+    return true;
+  }
+
+  utils::NMS(result, nms_iou_threshold);
+
+  // scale and clip box
+  for (size_t i = 0; i < result->boxes.size(); ++i) {
+    result->boxes[i][0] = std::max(result->boxes[i][0], 0.0f);
+    result->boxes[i][1] = std::max(result->boxes[i][1], 0.0f);
+    result->boxes[i][2] = std::max(result->boxes[i][2], 0.0f);
+    result->boxes[i][3] = std::max(result->boxes[i][3], 0.0f);
+    result->boxes[i][0] = std::min(result->boxes[i][0], ipt_w - 1.0f);
+    result->boxes[i][1] = std::min(result->boxes[i][1], ipt_h - 1.0f);
+    result->boxes[i][2] = std::min(result->boxes[i][2], ipt_w - 1.0f);
+    result->boxes[i][3] = std::min(result->boxes[i][3], ipt_h - 1.0f);
+  }
+  return true;
+}
+
+bool UltraFace::Predict(cv::Mat* im, FaceDetectionResult* result,
+                        float conf_threshold, float nms_iou_threshold) {
+#ifdef FASTDEPLOY_DEBUG
+  TIMERECORD_START(0)
+#endif
+
+  Mat mat(*im);
+  std::vector<FDTensor> input_tensors(1);
+
+  std::map<std::string, std::array<float, 2>> im_info;
+
+  // Record the shape of image and the shape of preprocessed image
+  im_info["input_shape"] = {static_cast<float>(mat.Height()),
+                            static_cast<float>(mat.Width())};
+  im_info["output_shape"] = {static_cast<float>(mat.Height()),
+                             static_cast<float>(mat.Width())};
+
+  if (!Preprocess(&mat, &input_tensors[0], &im_info)) {
+    FDERROR << "Failed to preprocess input image." << std::endl;
+    return false;
+  }
+
+#ifdef FASTDEPLOY_DEBUG
+  TIMERECORD_END(0, "Preprocess")
+  TIMERECORD_START(1)
+#endif
+
+  input_tensors[0].name = InputInfoOfRuntime(0).name;
+  std::vector<FDTensor> output_tensors;
+  if (!Infer(input_tensors, &output_tensors)) {
+    FDERROR << "Failed to inference." << std::endl;
+    return false;
+  }
+#ifdef FASTDEPLOY_DEBUG
+  TIMERECORD_END(1, "Inference")
+  TIMERECORD_START(2)
+#endif
+
+  if (!Postprocess(output_tensors, result, im_info, conf_threshold,
+                   nms_iou_threshold)) {
+    FDERROR << "Failed to post process." << std::endl;
+    return false;
+  }
+
+#ifdef FASTDEPLOY_DEBUG
+  TIMERECORD_END(2, "Postprocess")
+#endif
+  return true;
+}
+
+}  // namespace linzaer
+}  // namespace vision
+}  // namespace fastdeploy

csrcs/fastdeploy/vision/linzaer/ultraface.h

@@ -0,0 +1,84 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include "fastdeploy/fastdeploy_model.h"
+#include "fastdeploy/vision/common/processors/transform.h"
+#include "fastdeploy/vision/common/result.h"
+
+namespace fastdeploy {
+
+namespace vision {
+
+namespace linzaer {
+
+class FASTDEPLOY_DECL UltraFace : public FastDeployModel {
+ public:
+  // 当model_format为ONNX时，无需指定params_file
+  // 当model_format为Paddle时，则需同时指定model_file & params_file
+  UltraFace(const std::string& model_file, const std::string& params_file = "",
+            const RuntimeOption& custom_option = RuntimeOption(),
+            const Frontend& model_format = Frontend::ONNX);
+
+  // 定义模型的名称
+  std::string ModelName() const {
+    return "Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB";
+  }
+
+  // 模型预测接口，即用户调用的接口
+  // im 为用户的输入数据，目前对于CV均定义为cv::Mat
+  // result 为模型预测的输出结构体
+  // conf_threshold 为后处理的参数
+  // nms_iou_threshold 为后处理的参数
+  virtual bool Predict(cv::Mat* im, FaceDetectionResult* result,
+                       float conf_threshold = 0.7f,
+                       float nms_iou_threshold = 0.3f);
+
+  // 以下为模型在预测时的一些参数，基本是前后处理所需
+  // 用户在创建模型后，可根据模型的要求，以及自己的需求
+  // 对参数进行修改
+  // tuple of (width, height), default (320, 240)
+  std::vector<int> size;
+
+ private:
+  // 初始化函数，包括初始化后端，以及其它模型推理需要涉及的操作
+  bool Initialize();
+
+  // 输入图像预处理操作
+  // Mat为FastDeploy定义的数据结构
+  // FDTensor为预处理后的Tensor数据，传给后端进行推理
+  // im_info为预处理过程保存的数据，在后处理中需要用到
+  bool Preprocess(Mat* mat, FDTensor* outputs,
+                  std::map<std::string, std::array<float, 2>>* im_info);
+
+  // 后端推理结果后处理，输出给用户
+  // infer_result 为后端推理后的输出Tensor
+  // result 为模型预测的结果
+  // im_info 为预处理记录的信息，后处理用于还原box
+  // conf_threshold 后处理时过滤box的置信度阈值
+  // nms_iou_threshold 后处理时NMS设定的iou阈值
+  bool Postprocess(std::vector<FDTensor>& infer_result,
+                   FaceDetectionResult* result,
+                   const std::map<std::string, std::array<float, 2>>& im_info,
+                   float conf_threshold, float nms_iou_threshold);
+
+  // 查看输入是否为动态维度的 不建议直接使用 不同模型的逻辑可能不一致
+  bool IsDynamicInput() const { return is_dynamic_input_; }
+
+  bool is_dynamic_input_;
+};
+
+}  // namespace linzaer
+}  // namespace vision
+}  // namespace fastdeploy

csrcs/fastdeploy/vision/vision_pybind.cc

@@ -25,6 +25,7 @@ void BindMeituan(pybind11::module& m);
 void BindMegvii(pybind11::module& m);
 void BindDeepCam(pybind11::module& m);
 void BindRangiLyu(pybind11::module& m);
+void BindLinzaer(pybind11::module& m);
 #ifdef ENABLE_VISION_VISUALIZE
 void BindVisualize(pybind11::module& m);
 #endif
@@ -69,6 +70,7 @@ void BindVision(pybind11::module& m) {
   BindMegvii(m);
   BindDeepCam(m);
   BindRangiLyu(m);
+  BindLinzaer(m);
 #ifdef ENABLE_VISION_VISUALIZE
   BindVisualize(m);
 #endif

examples/vision/linzaer_ultraface.cc

@@ -0,0 +1,53 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision.h"
+
+int main() {
+  namespace vis = fastdeploy::vision;
+
+  std::string model_file = "../resources/models/version-RFB-320.onnx";
+  std::string img_path = "../resources/images/test_face_det_0.jpg";
+  std::string vis_path =
+      "../resources/outputs/linzaer_ultraface_vis_result.jpg";
+
+  auto model = vis::linzaer::UltraFace(model_file);
+  if (!model.Initialized()) {
+    std::cerr << "Init Failed! Model: " << model_file << std::endl;
+    return -1;
+  } else {
+    std::cout << "Init Done! Model:" << model_file << std::endl;
+  }
+  model.EnableDebug();
+
+  cv::Mat im = cv::imread(img_path);
+  cv::Mat vis_im = im.clone();
+
+  vis::FaceDetectionResult res;
+  if (!model.Predict(&im, &res, 0.7f, 0.3f)) {
+    std::cerr << "Prediction Failed." << std::endl;
+    return -1;
+  } else {
+    std::cout << "Prediction Done!" << std::endl;
+  }
+
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  // 可视化预测结果
+  vis::Visualize::VisFaceDetection(&vis_im, res, 2, 0.3f);
+  cv::imwrite(vis_path, vis_im);
+  std::cout << "Detect Done! Saved: " << vis_path << std::endl;
+  return 0;
+}

fastdeploy/vision/__init__.py

@@ -24,3 +24,4 @@ from . import visualize
 from . import wongkinyiu
 from . import deepcam
 from . import rangilyu
+from . import linzaer

fastdeploy/vision/linzaer/__init__.py

@@ -0,0 +1,53 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+import logging
+from ... import FastDeployModel, Frontend
+from ... import fastdeploy_main as C
+
+
+class UltraFace(FastDeployModel):
+    def __init__(self,
+                 model_file,
+                 params_file="",
+                 runtime_option=None,
+                 model_format=Frontend.ONNX):
+        # 调用基函数进行backend_option的初始化
+        # 初始化后的option保存在self._runtime_option
+        super(UltraFace, self).__init__(runtime_option)
+
+        self._model = C.vision.linzaer.UltraFace(
+            model_file, params_file, self._runtime_option, model_format)
+        # 通过self.initialized判断整个模型的初始化是否成功
+        assert self.initialized, "UltraFace initialize failed."
+
+    def predict(self, input_image, conf_threshold=0.7, nms_iou_threshold=0.3):
+        return self._model.predict(input_image, conf_threshold,
+                                   nms_iou_threshold)
+
+    # 一些跟UltraFace模型有关的属性封装
+    # 多数是预处理相关，可通过修改如model.size = [640, 480]改变预处理时resize的大小（前提是模型支持）
+    @property
+    def size(self):
+        return self._model.size
+
+    @size.setter
+    def size(self, wh):
+        assert isinstance(wh, [list, tuple]),\
+            "The value to set `size` must be type of tuple or list."
+        assert len(wh) == 2,\
+            "The value to set `size` must contatins 2 elements means [width, height], but now it contains {} elements.".format(
+            len(wh))
+        self._model.size = wh

model_zoo/vision/ultraface/README.md

@@ -0,0 +1,49 @@
+# UltraFace部署示例
+
+当前支持模型版本为：[UltraFace CommitID:dffdddd](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/commit/dffdddd)
+
+本文档说明如何进行[UltraFace](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/)的快速部署推理。本目录结构如下
+
+```
+.
+├── cpp                     # C++ 代码目录
+│   ├── CMakeLists.txt      # C++ 代码编译CMakeLists文件
+│   ├── README.md           # C++ 代码编译部署文档
+│   └── ultraface.cc        # C++ 示例代码
+├── api.md                  # API 说明文档
+├── README.md               # UltraFace 部署文档
+└── ultraface.py            # Python示例代码
+```
+
+## 安装FastDeploy
+
+使用如下命令安装FastDeploy，注意到此处安装的是`vision-cpu`，也可根据需求安装`vision-gpu`
+```bash
+# 安装fastdeploy-python工具
+pip install fastdeploy-python
+
+# 安装vision-cpu模块
+fastdeploy install vision-cpu
+```
+
+## Python部署
+
+执行如下代码即会自动下载YOLOv5Face模型和测试图片
+```bash
+python ultraface.py
+```
+
+执行完成后会将可视化结果保存在本地`vis_result.jpg`，同时输出检测结果如下
+```
+FaceDetectionResult: [xmin, ymin, xmax, ymax, score]
+742.528931,261.309937, 837.749146, 365.145599, 0.999833
+408.159332,253.410889, 484.747284, 353.378052, 0.999832
+549.409424,225.051819, 636.311890, 337.824707, 0.999782
+185.562805,233.364044, 252.001801, 323.948669, 0.999709
+304.065918,180.468140, 377.097961, 278.932861, 0.999645
+```
+
+## 其它文档
+
+- [C++部署](./cpp/README.md)
+- [UltraFace API文档](./api.md)

model_zoo/vision/ultraface/api.md

@@ -0,0 +1,71 @@
+# UltraFace API说明
+
+## Python API
+
+### UltraFace类
+```
+fastdeploy.vision.linzaer.UltraFace(model_file, params_file=None, runtime_option=None, model_format=fd.Frontend.ONNX)
+```
+UltraFace模型加载和初始化，当model_format为`fd.Frontend.ONNX`时，只需提供model_file，如`version-RFB-320.onnx`；当model_format为`fd.Frontend.PADDLE`时，则需同时提供model_file和params_file。
+
+**参数**
+
+> * **model_file**(str): 模型文件路径
+> * **params_file**(str): 参数文件路径
+> * **runtime_option**(RuntimeOption): 后端推理配置，默认为None，即采用默认配置
+> * **model_format**(Frontend): 模型格式
+
+#### predict函数
+> ```
+> UltraFace.predict(image_data, conf_threshold=0.7, nms_iou_threshold=0.3)
+> ```
+> 模型预测结口，输入图像直接输出检测结果。
+>
+> **参数**
+>
+> > * **image_data**(np.ndarray): 输入数据，注意需为HWC，BGR格式
+> > * **conf_threshold**(float): 检测框置信度过滤阈值
+> > * **nms_iou_threshold**(float): NMS处理过程中iou阈值
+
+示例代码参考[ultraface.py](./ultraface.py)
+
+
+## C++ API
+
+### UltraFace类
+```
+fastdeploy::vision::linzaer::UltraFace(
+        const string& model_file,
+        const string& params_file = "",
+        const RuntimeOption& runtime_option = RuntimeOption(),
+        const Frontend& model_format = Frontend::ONNX)
+```
+UltraFace模型加载和初始化，当model_format为`Frontend::ONNX`时，只需提供model_file，如`version-RFB-320.onnx`；当model_format为`Frontend::PADDLE`时，则需同时提供model_file和params_file。
+
+**参数**
+
+> * **model_file**(str): 模型文件路径
+> * **params_file**(str): 参数文件路径
+> * **runtime_option**(RuntimeOption): 后端推理配置，默认为None，即采用默认配置
+> * **model_format**(Frontend): 模型格式
+
+#### Predict函数
+> ```
+> UltraFace::Predict(cv::Mat* im, FaceDetectionResult* result,
+>                    float conf_threshold = 0.7,
+>                    float nms_iou_threshold = 0.3)
+> ```
+> 模型预测接口，输入图像直接输出检测结果。
+>
+> **参数**
+>
+> > * **im**: 输入图像，注意需为HWC，BGR格式
+> > * **result**: 检测结果，包括检测框，各个框的置信度
+> > * **conf_threshold**: 检测框置信度过滤阈值
+> > * **nms_iou_threshold**: NMS处理过程中iou阈值
+
+示例代码参考[cpp/ultraface.cc](cpp/ultraface.cc)
+
+## 其它API使用
+
+- [模型部署RuntimeOption配置](../../../docs/api/runtime_option.md)

model_zoo/vision/ultraface/cpp/CMakeLists.txt

@@ -0,0 +1,17 @@
+PROJECT(ultraface_demo C CXX)
+CMAKE_MINIMUM_REQUIRED (VERSION 3.16)
+
+# 在低版本ABI环境中，通过如下代码进行兼容性编译
+# add_definitions(-D_GLIBCXX_USE_CXX11_ABI=0)
+
+# 指定下载解压后的fastdeploy库路径
+set(FASTDEPLOY_INSTALL_DIR ${PROJECT_SOURCE_DIR}/fastdeploy-linux-x64-0.3.0/)
+
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake)
+
+# 添加FastDeploy依赖头文件
+include_directories(${FASTDEPLOY_INCS})
+
+add_executable(ultraface_demo ${PROJECT_SOURCE_DIR}/ultraface.cc)
+# 添加FastDeploy库依赖
+target_link_libraries(ultraface_demo ${FASTDEPLOY_LIBS})

model_zoo/vision/ultraface/cpp/README.md

@@ -0,0 +1,36 @@
+# 编译UltraFace示例
+
+当前支持模型版本为：[UltraFace CommitID:dffdddd](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/commit/dffdddd)
+
+## 下载和解压预测库
+```bash
+wget https://bj.bcebos.com/paddle2onnx/fastdeploy/fastdeploy-linux-x64-0.0.3.tgz
+tar xvf fastdeploy-linux-x64-0.0.3.tgz
+```
+
+## 编译示例代码
+```bash
+mkdir build & cd build
+cmake ..
+make -j
+```
+
+## 下载模型和图片
+wget https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/raw/master/models/onnx/version-RFB-320.onnx  
+wget https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/raw/master/imgs/3.jpg
+
+
+## 执行
+```bash
+./ultraface_demo
+```
+
+执行完后可视化的结果保存在本地`vis_result.jpg`，同时会将检测框输出在终端，如下所示
+```
+FaceDetectionResult: [xmin, ymin, xmax, ymax, score]
+742.528931,261.309937, 837.749146, 365.145599, 0.999833
+408.159332,253.410889, 484.747284, 353.378052, 0.999832
+549.409424,225.051819, 636.311890, 337.824707, 0.999782
+185.562805,233.364044, 252.001801, 323.948669, 0.999709
+304.065918,180.468140, 377.097961, 278.932861, 0.999645
+```

model_zoo/vision/ultraface/cpp/ultraface.cc

@@ -0,0 +1,48 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision.h"
+
+int main() {
+  namespace vis = fastdeploy::vision;
+
+  auto model = vis::linzaer::UltraFace("version-RFB-320.onnx");
+  if (!model.Initialized()) {
+    std::cerr << "Init Failed! Model: " << model_file << std::endl;
+    return -1;
+  } else {
+    std::cout << "Init Done! Model:" << model_file << std::endl;
+  }
+  model.EnableDebug();
+
+  cv::Mat im = cv::imread("3.jpg");
+  cv::Mat vis_im = im.clone();
+
+  vis::FaceDetectionResult res;
+  if (!model.Predict(&im, &res, 0.7f, 0.3f)) {
+    std::cerr << "Prediction Failed." << std::endl;
+    return -1;
+  } else {
+    std::cout << "Prediction Done!" << std::endl;
+  }
+
+  // 输出预测框结果
+  std::cout << res.Str() << std::endl;
+
+  // 可视化预测结果
+  vis::Visualize::VisFaceDetection(&vis_im, res, 2, 0.3f);
+  cv::imwrite("vis_result.jpg", vis_im);
+  std::cout << "Detect Done! Saved: " << vis_path << std::endl;
+  return 0;
+}

model_zoo/vision/ultraface/ultraface.py

@@ -0,0 +1,23 @@
+import fastdeploy as fd
+import cv2
+
+# 下载模型
+model_url = "https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/raw/master/models/onnx/version-RFB-320.onnx"
+test_img_url = "https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/raw/master/imgs/3.jpg"
+fd.download(model_url, ".", show_progress=True)
+fd.download(test_img_url, ".", show_progress=True)
+
+# 加载模型
+model = fd.vision.linzaer.UltraFace("version-RFB-320.onnx")
+
+# 预测图片
+im = cv2.imread("3.jpg")
+result = model.predict(im, conf_threshold=0.7, nms_iou_threshold=0.3)
+
+# 可视化结果
+fd.vision.visualize.vis_face_detection(im, result)
+cv2.imwrite("vis_result.jpg", im)
+
+# 输出预测结果
+print(result)
+print(model.runtime_option)

model_zoo/vision/yolov5face/cpp/README.md

@@ -50,7 +50,7 @@ make -j
 
 执行完后可视化的结果保存在本地`vis_result.jpg`，同时会将检测框输出在终端，如下所示
 ```
-aceDetectionResult: [xmin, ymin, xmax, ymax, score, (x, y) x 5]
+FaceDetectionResult: [xmin, ymin, xmax, ymax, score, (x, y) x 5]
 749.575256,375.122162, 775.008850, 407.858215, 0.851824, (756.933838,388.423157), (767.810974,387.932922), (762.617065,394.212341), (758.053101,399.073639), (767.370300,398.769470)
 897.833862,380.372864, 924.725281, 409.566803, 0.847505, (903.757202,390.221741), (914.575867,389.495911), (908.998901,395.983307), (905.803223,400.871429), (914.674438,400.268066)
 281.558197,367.739349, 305.474701, 397.860535, 0.840915, (287.018768,379.771088), (297.285004,378.755280), (292.057831,385.207367), (289.110962,390.010437), (297.535339,389.412048)