diff --git a/deploy/python/README.md b/deploy/python/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b9a8cc35eb325853dba329b0b1c75c44656b016b
--- /dev/null
+++ b/deploy/python/README.md
@@ -0,0 +1,72 @@
+## PaddleDetection Python 预测部署方案
+本篇教程使用AnalysisPredictor对[导出模型](../../docs/advanced_tutorials/inference/EXPORT_MODEL.md)进行高性能预测。
+
+在PaddlePaddle中预测引擎和训练引擎底层有着不同的优化方法, 下面列出了两种不同的预测方式。Executor同时支持训练和预测，AnalysisPredictor则专门针对推理进行了优化，是基于[C++预测库](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/native_infer.html)的Python接口，该引擎可以对模型进行多项图优化，减少不必要的内存拷贝。如果用户在部署已训练模型的过程中对性能有较高的要求，于是我们提供了独立于PaddleDetection的预测脚本，方便用户直接集成部署。
+
+- Executor：[Executor](https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/basic_concept/executor.html#executor)
+- AnalysisPredictor：[AnalysisPredictor](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/python_infer_cn.html#analysispredictor)
+
+
+主要包含两个步骤：
+
+- 导出预测模型
+- 基于Python的预测
+
+## 1. 导出预测模型
+
+PaddleDetection在训练过程包括网络的前向和优化器相关参数，而在部署过程中，我们只需要前向参数，具体参考:[导出模型](../../docs/advanced_tutorials/inference/EXPORT_MODEL.md)
+
+导出后目录下，包括`__model__`，`__params__`和`infer_cfg.yml`三个文件。
+
+## 2. 基于python的预测
+### 2.1 安装依赖
+  - `PaddlePaddle`的安装:
+    请点击[官方安装文档](https://paddlepaddle.org.cn/install/quick) 选择适合的方式，版本为1.7以上即可
+  - 切换到`PaddleDetection`代码库根目录，执行`pip install -r requirements.txt`安装其它依赖
+
+### 2.2 执行预测程序
+在终端输入以下命令进行预测：
+
+```bash
+python infer.py --models_dir=/path/to/models --image_file=/path/to/image
+--use_gpu=(False/True)
+```
+
+参数说明如下:
+
+| 参数 | 是否必须|含义 |
+|-------|-------|----------|
+| --models_dir | Yes|上述导出的模型路径 |
+| --image_file | Yes |需要预测的图片 |
+| --video_file | Yes |需要预测的视频 |
+| --use_gpu |No|是否GPU，默认为False|
+| --threshold |No|预测得分的阈值，默认为0.5|
+| --visualize |No|是否可视化结果，默认为False|
+| --output_dir |No|可视化结果保存的根目录，默认为output/|
+
+
+## 3. 部署性能对比测试
+对比AnalysisPredictor相对Executor的推理速度
+
+### 3.1 测试环境:
+
+- CUDA 9.0
+- CUDNN 7.5
+- PaddlePaddle 1.71
+- GPU: Tesla P40
+
+### 3.2 测试方式:
+
+- Batch Size=1
+- 去掉前100轮warmup时间，测试100轮的平均时间，单位ms/image，只计算模型运行时间，不包括数据的处理和拷贝。
+
+
+### 3.3 测试结果
+
+|模型 | AnalysisPredictor | Executor | 输入|
+|---|----|---|---|
+| YOLOv3-MobileNetv1 | 15.20 | 19.54 |  608*608
+| faster_rcnn_r50_fpn_1x | 50.05 | 69.58 |800*1088
+| faster_rcnn_r50_1x | 326.11 | 347.22 | 800*1067
+| mask_rcnn_r50_fpn_1x | 67.49 | 91.02 | 800*1088
+| mask_rcnn_r50_1x | 326.11 | 350.94 | 800*1067
diff --git a/deploy/python/infer.py b/deploy/python/infer.py
new file mode 100644
index 0000000000000000000000000000000000000000..8e7a261f24406ef2f8b64b698bd98a3ba70203ce
--- /dev/null
+++ b/deploy/python/infer.py
@@ -0,0 +1,556 @@
+# Copyright (c) 2020 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.
+
+import os
+import argparse
+import yaml
+from PIL import Image
+import cv2
+import numpy as np
+import paddle.fluid as fluid
+from visualize import visualize_box_mask
+
+
+def decode_image(im_file, im_info):
+    """read rgb image
+    Args:
+        im_file (str/np.ndarray): path of image/ np.ndarray read by cv2
+        im_info (dict): info of image
+    Returns:
+        im (np.ndarray):  processed image (np.ndarray)
+        im_info (dict): info of processed image
+    """
+    if isinstance(im_file, str):
+        with open(im_file, 'rb') as f:
+            im_read = f.read()
+        data = np.frombuffer(im_read, dtype='uint8')
+        im = cv2.imdecode(data, 1)  # BGR mode, but need RGB mode
+        im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+        im_info['origin_shape'] = im.shape[:2]
+        im_info['resize_shape'] = im.shape[:2]
+    else:
+        im = im_file
+        im_info['origin_shape'] = im.shape[:2]
+        im_info['resize_shape'] = im.shape[:2]
+    return im, im_info
+
+
+class Resize(object):
+    """resize image by target_size and max_size
+    Args:
+        arch (str): model type
+        target_size (int): the target size of image
+        max_size (int): the max size of image
+        use_cv2 (bool): whether us cv2
+        image_shape (list): input shape of model
+        interp (int): method of resize
+    """
+
+    def __init__(self,
+                 arch,
+                 target_size,
+                 max_size,
+                 use_cv2=True,
+                 image_shape=None,
+                 interp=cv2.INTER_LINEAR):
+        self.target_size = target_size
+        self.max_size = max_size
+        self.image_shape = image_shape,
+        self.arch = arch
+        self.use_cv2 = use_cv2
+        self.interp = interp
+        self.scale_set = {'RCNN', 'RetinaNet'}
+
+    def __call__(self, im, im_info):
+        """
+        Args:
+            im (np.ndarray): image (np.ndarray)
+            im_info (dict): info of image
+        Returns:
+            im (np.ndarray):  processed image (np.ndarray)
+            im_info (dict): info of processed image
+        """
+        im_channel = im.shape[2]
+        im_scale_x, im_scale_y = self.generate_scale(im)
+        if self.use_cv2:
+            im = cv2.resize(
+                im,
+                None,
+                None,
+                fx=im_scale_x,
+                fy=im_scale_y,
+                interpolation=self.interp)
+        else:
+            resize_w = int(im_scale_x * float(im.shape[1]))
+            resize_h = int(im_scale_y * float(im.shape[0]))
+            if self.max_size != 0:
+                raise TypeError(
+                    'If you set max_size to cap the maximum size of image,'
+                    'please set use_cv2 to True to resize the image.')
+            im = im.astype('uint8')
+            im = Image.fromarray(im)
+            im = im.resize((int(resize_w), int(resize_h)), self.interp)
+            im = np.array(im)
+
+        # padding im when image_shape fixed by infer_cfg.yml
+        if self.max_size != 0 and self.image_shape is not None:
+            padding_im = np.zeros(
+                (self.max_size, self.max_size, im_channel), dtype=np.float32)
+            im_h, im_w = im.shape[:2]
+            padding_im[:im_h, :im_w, :] = im
+            im = padding_im
+
+        if self.arch in self.scale_set:
+            im_info['scale'] = im_scale_x
+        im_info['resize_shape'] = im.shape[:2]
+        return im, im_info
+
+    def generate_scale(self, im):
+        """
+        Args:
+            im (np.ndarray): image (np.ndarray)
+        Returns:
+            im_scale_x: the resize ratio of X 
+            im_scale_y: the resize ratio of Y 
+        """
+        origin_shape = im.shape[:2]
+        im_c = im.shape[2]
+        if self.max_size != 0 and self.arch in self.scale_set:
+            im_size_min = np.min(origin_shape[0:2])
+            im_size_max = np.max(origin_shape[0:2])
+            im_scale = float(self.target_size) / float(im_size_min)
+            if np.round(im_scale * im_size_max) > self.max_size:
+                im_scale = float(self.max_size) / float(im_size_max)
+            im_scale_x = im_scale
+            im_scale_y = im_scale
+        else:
+            im_scale_x = float(self.target_size) / float(origin_shape[1])
+            im_scale_y = float(self.target_size) / float(origin_shape[0])
+        return im_scale_x, im_scale_y
+
+
+class Normalize(object):
+    """normalize image
+    Args:
+        mean (list): im - mean
+        std (list): im / std
+        is_scale (bool): whether need im / 255
+        is_channel_first (bool): if True: image shape is CHW, else: HWC
+    """
+
+    def __init__(self, mean, std, is_scale=True, is_channel_first=False):
+        self.mean = mean
+        self.std = std
+        self.is_scale = is_scale
+        self.is_channel_first = is_channel_first
+
+    def __call__(self, im, im_info):
+        """
+        Args:
+            im (np.ndarray): image (np.ndarray)
+            im_info (dict): info of image
+        Returns:
+            im (np.ndarray):  processed image (np.ndarray)
+            im_info (dict): info of processed image
+        """
+        im = im.astype(np.float32, copy=False)
+        if self.is_channel_first:
+            mean = np.array(self.mean)[:, np.newaxis, np.newaxis]
+            std = np.array(self.std)[:, np.newaxis, np.newaxis]
+        else:
+            mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
+            std = np.array(self.std)[np.newaxis, np.newaxis, :]
+        if self.is_scale:
+            im = im / 255.0
+        im -= mean
+        im /= std
+        return im, im_info
+
+
+class Permute(object):
+    """permute image
+    Args:
+        to_bgr (bool): whether convert RGB to BGR 
+        channel_first (bool): whether convert HWC to CHW
+    """
+
+    def __init__(self, to_bgr=False, channel_first=True):
+        self.to_bgr = to_bgr
+        self.channel_first = channel_first
+
+    def __call__(self, im, im_info):
+        """
+        Args:
+            im (np.ndarray): image (np.ndarray)
+            im_info (dict): info of image
+        Returns:
+            im (np.ndarray):  processed image (np.ndarray)
+            im_info (dict): info of processed image
+        """
+        if self.channel_first:
+            im = im.transpose((2, 0, 1)).copy()
+        if self.to_bgr:
+            im = im[[2, 1, 0], :, :]
+        return im, im_info
+
+
+class PadStride(object):
+    """ padding image for model with FPN 
+    Args:
+        stride (bool): model with FPN need image shape % stride == 0 
+    """
+
+    def __init__(self, stride=0):
+        self.coarsest_stride = stride
+
+    def __call__(self, im, im_info):
+        """
+        Args:
+            im (np.ndarray): image (np.ndarray)
+            im_info (dict): info of image
+        Returns:
+            im (np.ndarray):  processed image (np.ndarray)
+            im_info (dict): info of processed image
+        """
+        coarsest_stride = self.coarsest_stride
+        if coarsest_stride == 0:
+            return im
+        im_c, im_h, im_w = im.shape
+        pad_h = int(np.ceil(float(im_h) / coarsest_stride) * coarsest_stride)
+        pad_w = int(np.ceil(float(im_w) / coarsest_stride) * coarsest_stride)
+        padding_im = np.zeros((im_c, pad_h, pad_w), dtype=np.float32)
+        padding_im[:, :im_h, :im_w] = im
+        im_info['resize_shape'] = padding_im.shape[1:]
+        return padding_im, im_info
+
+
+def create_inputs(im, im_info, model_arch='YOLO'):
+    """generate input for different model type
+    Args:
+        im (np.ndarray): image (np.ndarray)
+        im_info (dict): info of image
+        model_arch (str): model type
+    Returns:
+        inputs (dict): input of model
+    """
+    inputs = {}
+    inputs['image'] = im
+    origin_shape = list(im_info['origin_shape'])
+    resize_shape = list(im_info['resize_shape'])
+    scale = im_info['scale']
+    if 'YOLO' in model_arch:
+        im_size = np.array([origin_shape]).astype('int32')
+        inputs['im_size'] = im_size
+    elif 'RetinaNet' in model_arch:
+        im_info = np.array([resize_shape + [scale]]).astype('float32')
+        inputs['im_info'] = im_info
+    elif 'RCNN' in model_arch:
+        im_info = np.array([resize_shape + [scale]]).astype('float32')
+        im_shape = np.array([origin_shape + [1.]]).astype('float32')
+        inputs['im_info'] = im_info
+        inputs['im_shape'] = im_shape
+    return inputs
+
+
+class Config():
+    """set config of preprocess, postprocess and visualize
+    Args:
+        model_dir (str): root path of model.yml
+    """
+    support_models = ['YOLO', 'SSD', 'RetinaNet', 'RCNN']
+
+    def __init__(self, model_dir):
+        # parsing Yaml config for Preprocess
+        deploy_file = os.path.join(model_dir, 'infer_cfg.yml')
+        with open(deploy_file) as f:
+            yml_conf = yaml.safe_load(f)
+        self.check_model(yml_conf)
+        self.arch = yml_conf['arch']
+        self.preprocess_infos = yml_conf['Preprocess']
+        self.use_python_inference = yml_conf['use_python_inference']
+        self.run_mode = yml_conf['mode']
+        self.min_subgraph_size = yml_conf['min_subgraph_size']
+        self.labels = yml_conf['label_list']
+        if not yml_conf['with_background']:
+            self.labels = self.labels[1:]
+        self.mask_resolution = None
+        if 'mask_resolution' in yml_conf:
+            self.mask_resolution = yml_conf['mask_resolution']
+
+    def check_model(self, yml_conf):
+        """
+        Raises:
+            ValueError: loaded model not in supported model type 
+        """
+        for support_model in self.support_models:
+            if support_model in yml_conf['arch']:
+                return True
+        raise ValueError(
+            "Unsupported arch: {}, expect SSD, YOLO, RetinaNet and RCNN".format(
+                yml_conf['arch']))
+
+
+def load_predictor(model_dir,
+                   run_mode='fluid',
+                   batch_size=1,
+                   use_gpu=False,
+                   min_subgraph_size=3):
+    """set AnalysisConfig，generate AnalysisPredictor
+    Args:
+        model_dir (str): root path of __model__ and __params__
+        use_gpu (bool): whether use gpu
+    Returns:
+        predictor (PaddlePredictor): AnalysisPredictor
+    Raises:
+        ValueError: predict by TensorRT need use_gpu == True 
+    """
+    if not use_gpu and run_mode == 'fluid':
+        raise ValueError(
+            "Predict by TensorRT mode: {}, expect use_gpu==True, but use_gpu == {}"
+            .format(run_mode, use_gpu))
+    precision_map = {
+        'trt_int8': fluid.core.AnalysisConfig.Precision.Int8,
+        'trt_fp32': fluid.core.AnalysisConfig.Precision.Float32,
+        'trt_fp16': fluid.core.AnalysisConfig.Precision.Half
+    }
+    config = fluid.core.AnalysisConfig(
+        os.path.join(model_dir, '__model__'),
+        os.path.join(model_dir, '__params__'))
+    if use_gpu:
+        # initial GPU memory(M), device ID
+        config.enable_use_gpu(100, 0)
+        # optimize graph and fuse op
+        config.switch_ir_optim(True)
+    else:
+        config.disable_gpu()
+
+    if run_mode in precision_map.keys():
+        config.enable_tensorrt_engine(
+            workspace_size=1 << 30,
+            max_batch_size=batch_size,
+            min_subgraph_size=min_subgraph_size,
+            precision_mode=precision_map[run_mode],
+            use_static=False,
+            use_calib_mode=run_mode == 'trt_int8')
+
+    # disable print log when predict
+    config.disable_glog_info()
+    # enable shared memory
+    config.enable_memory_optim()
+    # disable feed, fetch OP，needed by zero_copy_run
+    config.switch_use_feed_fetch_ops(False)
+    predictor = fluid.core.create_paddle_predictor(config)
+    return predictor
+
+
+def load_executor(model_dir, use_gpu=False):
+    if use_gpu:
+        place = fluid.CUDAPlace(0)
+    else:
+        place = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    program, feed_names, fetch_targets = fluid.io.load_inference_model(
+        dirname=model_dir,
+        executor=exe,
+        model_filename='__model__',
+        params_filename='__params__')
+    return exe, program, fetch_targets
+
+
+def visualize(image_file,
+              results,
+              labels,
+              mask_resolution=14,
+              output_dir='output/'):
+    # visualize the predict result
+    im = visualize_box_mask(
+        image_file, results, labels, mask_resolution=mask_resolution)
+    img_name = os.path.split(image_file)[-1]
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+    out_path = os.path.join(output_dir, img_name)
+    im.save(out_path, quality=95)
+    print("save result to: " + out_path)
+
+
+class Detector():
+    """
+    Args:
+        model_dir (str): root path of __model__, __params__ and infer_cfg.yml
+        use_gpu (bool): whether use gpu
+    """
+
+    def __init__(self, model_dir, use_gpu=False, threshold=0.5):
+        self.config = Config(model_dir)
+        if self.config.use_python_inference:
+            self.executor, self.program, self.fecth_targets = load_executor(
+                model_dir, use_gpu=use_gpu)
+        else:
+            self.predictor = load_predictor(
+                model_dir,
+                run_mode=self.config.run_mode,
+                min_subgraph_size=self.config.min_subgraph_size,
+                use_gpu=use_gpu)
+        self.preprocess_ops = []
+        for op_info in self.config.preprocess_infos:
+            op_type = op_info.pop('type')
+            if op_type == 'Resize':
+                op_info['arch'] = self.config.arch
+            self.preprocess_ops.append(eval(op_type)(**op_info))
+
+    def preprocess(self, im):
+        # process image by preprocess_ops
+        im_info = {
+            'scale': 1.,
+            'origin_shape': None,
+            'resize_shape': None,
+        }
+        im, im_info = decode_image(im, im_info)
+        for operator in self.preprocess_ops:
+            im, im_info = operator(im, im_info)
+        im = np.array((im, )).astype('float32')
+        inputs = create_inputs(im, im_info, self.config.arch)
+        return inputs, im_info
+
+    def postprocess(self, np_boxes, np_masks, im_info, threshold=0.5):
+        # postprocess output of predictor
+        results = {}
+        if 'SSD' in self.config.arch:
+            w, h = im_info['origin_shape']
+            np_boxes[:, 2] *= h
+            np_boxes[:, 3] *= w
+            np_boxes[:, 4] *= h
+            np_boxes[:, 5] *= w
+        expect_boxes = np_boxes[:, 1] > threshold
+        np_boxes = np_boxes[expect_boxes, :]
+        for box in np_boxes:
+            print('class_id:{:d}, confidence:{:.2f},'
+                  'left_top:[{:.2f},{:.2f}],'
+                  ' right_bottom:[{:.2f},{:.2f}]'.format(
+                      int(box[0]), box[1], box[2], box[3], box[4], box[5]))
+        results['boxes'] = np_boxes
+        if np_masks is not None:
+            np_masks = np_masks[expect_boxes, :, :, :]
+            results['masks'] = np_masks
+        return results
+
+    def predict(self, image, threshold=0.5):
+        '''
+        Args:
+            image (str/np.ndarray): path of image/ np.ndarray read by cv2
+            threshold (float): threshold of predicted box' score
+        Returns:
+            results (dict): include 'boxes': np.ndarray: shape:[N,6], N: number of box，
+                            matix element:[class, score, x_min, y_min, x_max, y_max]
+                            MaskRCNN's results include 'masks': np.ndarray: 
+                            shape:[N, class_num, mask_resolution, mask_resolution]  
+        '''
+        inputs, im_info = self.preprocess(image)
+        np_boxes, np_masks = None, None
+        if self.config.use_python_inference:
+            outs = self.executor.run(self.program,
+                                     feed=inputs,
+                                     fetch_list=self.fecth_targets,
+                                     return_numpy=False)
+            np_boxes = np.array(outs[0])
+            if self.config.mask_resolution is not None:
+                np_masks = np.arrya(outs[1])
+        else:
+            input_names = self.predictor.get_input_names()
+            for i in range(len(inputs)):
+                input_tensor = self.predictor.get_input_tensor(input_names[i])
+                input_tensor.copy_from_cpu(inputs[input_names[i]])
+            self.predictor.zero_copy_run()
+            output_names = self.predictor.get_output_names()
+            boxes_tensor = self.predictor.get_output_tensor(output_names[0])
+            np_boxes = boxes_tensor.copy_to_cpu()
+            if self.config.mask_resolution is not None:
+                masks_tensor = self.predictor.get_output_tensor(output_names[1])
+                np_masks = masks_tensor.copy_to_cpu()
+        results = self.postprocess(
+            np_boxes, np_masks, im_info, threshold=threshold)
+        return results
+
+
+def predict_image():
+    detector = Detector(FLAGS.model_dir, use_gpu=FLAGS.use_gpu)
+    results = detector.predict(FLAGS.image_file, FLAGS.threshold)
+    visualize(
+        FLAGS.image_file,
+        results,
+        detector.config.labels,
+        mask_resolution=detector.config.mask_resolution,
+        output_dir=FLAGS.output_dir)
+
+
+def predict_video():
+    detector = Detector(FLAGS.model_dir, use_gpu=FLAGS.use_gpu)
+    capture = cv2.VideoCapture(FLAGS.video_file)
+    fps = 30
+    width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
+    video_name = os.path.split(FLAGS.video_file)[-1]
+    if not os.path.exists(FLAGS.output_dir):
+        os.makedirs(FLAGES.output_dir)
+    out_path = os.path.join(FLAGS.output_dir, video_name)
+    writer = cv2.VideoWriter(out_path, fourcc, fps, (width, height))
+    index = 1
+    while (1):
+        ret, frame = capture.read()
+        if not ret:
+            break
+        print('detect frame:%d' % (index))
+        index += 1
+        results = detector.predict(frame, FLAGS.threshold)
+        im = visualize_box_mask(
+            frame,
+            results,
+            detector.config.labels,
+            mask_resolution=detector.config.mask_resolution)
+        im = np.array(im)
+        writer.write(im)
+    writer.release()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--model_dir",
+        type=str,
+        default=None,
+        help=("Directory include:'__model__', '__params__', "
+              "'infer_cfg.yml', created by export_model."),
+        required=True)
+    parser.add_argument(
+        "--image_file", type=str, default='', help="Path of image file.")
+    parser.add_argument(
+        "--video_file", type=str, default='', help="Path of video file.")
+    parser.add_argument(
+        "--use_gpu", default=False, help="Whether to predict with GPU.")
+    parser.add_argument(
+        "--threshold", type=float, default=0.5, help="Threshold of score.")
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="output",
+        help="Directory of output visualization files.")
+
+    FLAGS = parser.parse_args()
+    if FLAGS.image_file != '' and FLAGS.video_file != '':
+        assert "Cannot predict image and video at the same time"
+    if FLAGS.image_file != '':
+        predict_image()
+    if FLAGS.video_file != '':
+        predict_video()
diff --git a/deploy/python/visualize.py b/deploy/python/visualize.py
new file mode 100644
index 0000000000000000000000000000000000000000..f58bee315010f03b39674e7df453c0b654244ae7
--- /dev/null
+++ b/deploy/python/visualize.py
@@ -0,0 +1,186 @@
+# coding: utf-8
+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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.
+
+import cv2
+import numpy as np
+from PIL import Image, ImageDraw
+
+
+def visualize_box_mask(im, results, labels, mask_resolution=14):
+    """ 
+    Args:
+        im (str/np.ndarray): path of image/np.ndarray read by cv2
+        results (dict): include 'boxes': np.ndarray: shape:[N,6], N: number of box，
+                        matix element:[class, score, x_min, y_min, x_max, y_max]
+                        MaskRCNN's results include 'masks': np.ndarray: 
+                        shape:[N, class_num, mask_resolution, mask_resolution]  
+        labels (list): labels:['class1', ..., 'classn']
+        mask_resolution (int): shape of a mask is:[mask_resolution, mask_resolution]
+    Returns:
+        im (PIL.Image.Image): visualized image  
+    """
+    if im is str:
+        im = Image.open(im).convert('RGB')
+    else:
+        im = Image.fromarray(im)
+    if 'masks' in results and 'boxes' in results:
+        im = draw_mask(
+            im,
+            results['boxes'],
+            results['masks'],
+            labels,
+            resolution=mask_resolution)
+    if 'boxes' in results:
+        im = draw_box(im, results['boxes'], labels)
+    return im
+
+
+def get_color_map_list(num_classes):
+    """ 
+    Args:
+        num_classes (int): number of class
+    Returns:
+        color_map (list): RGB color list 
+    """
+    color_map = num_classes * [0, 0, 0]
+    for i in range(0, num_classes):
+        j = 0
+        lab = i
+        while lab:
+            color_map[i * 3] |= (((lab >> 0) & 1) << (7 - j))
+            color_map[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j))
+            color_map[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j))
+            j += 1
+            lab >>= 3
+    color_map = [color_map[i:i + 3] for i in range(0, len(color_map), 3)]
+    return color_map
+
+
+def expand_boxes(boxes, scale=0.0):
+    """ 
+    Args:
+        boxes (np.ndarray): shape:[N,4], N:number of box，
+                            matix element:[x_min, y_min, x_max, y_max]
+        scale (float): scale of boxes
+    Returns:
+        boxes_exp (np.ndarray): expanded boxes
+    """
+    w_half = (boxes[:, 2] - boxes[:, 0]) * .5
+    h_half = (boxes[:, 3] - boxes[:, 1]) * .5
+    x_c = (boxes[:, 2] + boxes[:, 0]) * .5
+    y_c = (boxes[:, 3] + boxes[:, 1]) * .5
+    w_half *= scale
+    h_half *= scale
+    boxes_exp = np.zeros(boxes.shape)
+    boxes_exp[:, 0] = x_c - w_half
+    boxes_exp[:, 2] = x_c + w_half
+    boxes_exp[:, 1] = y_c - h_half
+    boxes_exp[:, 3] = y_c + h_half
+    return boxes_exp
+
+
+def draw_mask(im, np_boxes, np_masks, labels, resolution=14, threshold=0.5):
+    """ 
+    Args:
+        im (PIL.Image.Image): PIL image
+        np_boxes (np.ndarray): shape:[N,6], N: number of box，
+                               matix element:[class, score, x_min, y_min, x_max, y_max]
+        np_masks (np.ndarray): shape:[N, class_num, resolution, resolution]
+        labels (list): labels:['class1', ..., 'classn']
+        resolution (int): shape of a mask is:[resolution, resolution]
+        threshold (float): threshold of mask
+    Returns:
+        im (PIL.Image.Image): visualized image  
+    """
+    color_list = get_color_map_list(len(labels))
+    scale = (resolution + 2.0) / resolution
+    im_w, im_h = im.size
+    w_ratio = 0.4
+    alpha = 0.7
+    im = np.array(im).astype('float32')
+    rects = np_boxes[:, 2:]
+    expand_rects = expand_boxes(rects, scale)
+    expand_rects = expand_rects.astype(np.int32)
+    clsid_scores = np_boxes[:, 0:2]
+    padded_mask = np.zeros((resolution + 2, resolution + 2), dtype=np.float32)
+    clsid2color = {}
+    for idx in range(len(np_boxes)):
+        clsid, score = clsid_scores[idx].tolist()
+        clsid = int(clsid)
+        xmin, ymin, xmax, ymax = expand_rects[idx].tolist()
+        w = xmax - xmin + 1
+        h = ymax - ymin + 1
+        w = np.maximum(w, 1)
+        h = np.maximum(h, 1)
+        padded_mask[1:-1, 1:-1] = np_masks[idx, int(clsid), :, :]
+        resized_mask = cv2.resize(padded_mask, (w, h))
+        resized_mask = np.array(resized_mask > threshold, dtype=np.uint8)
+        x0 = min(max(xmin, 0), im_w)
+        x1 = min(max(xmax + 1, 0), im_w)
+        y0 = min(max(ymin, 0), im_h)
+        y1 = min(max(ymax + 1, 0), im_h)
+        im_mask = np.zeros((im_h, im_w), dtype=np.uint8)
+        im_mask[y0:y1, x0:x1] = resized_mask[(y0 - ymin):(y1 - ymin), (
+            x0 - xmin):(x1 - xmin)]
+        if clsid not in clsid2color:
+            clsid2color[clsid] = color_list[clsid]
+        color_mask = clsid2color[clsid]
+        for c in range(3):
+            color_mask[c] = color_mask[c] * (1 - w_ratio) + w_ratio * 255
+        idx = np.nonzero(im_mask)
+        color_mask = np.array(color_mask)
+        im[idx[0], idx[1], :] *= 1.0 - alpha
+        im[idx[0], idx[1], :] += alpha * color_mask
+    return Image.fromarray(im.astype('uint8'))
+
+
+def draw_box(im, np_boxes, labels):
+    """ 
+    Args:
+        im (PIL.Image.Image): PIL image
+        np_boxes (np.ndarray): shape:[N,6], N: number of box，
+                               matix element:[class, score, x_min, y_min, x_max, y_max]
+        labels (list): labels:['class1', ..., 'classn']
+    Returns:
+        im (PIL.Image.Image): visualized image  
+    """
+    draw_thickness = min(im.size) / 320
+    draw = ImageDraw.Draw(im)
+    clsid2color = {}
+    color_list = get_color_map_list(len(labels))
+
+    for dt in np_boxes:
+        clsid, bbox, score = int(dt[0]), dt[2:], dt[1]
+        xmin, ymin, xmax, ymax = bbox
+        w = xmax - xmin
+        h = ymax - ymin
+        if clsid not in clsid2color:
+            clsid2color[clsid] = color_list[clsid]
+        color = tuple(clsid2color[clsid])
+
+        # draw bbox
+        draw.line(
+            [(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin),
+             (xmin, ymin)],
+            width=draw_thickness,
+            fill=color)
+
+        # draw label
+        text = "{} {:.2f}".format(labels[clsid], score)
+        tw, th = draw.textsize(text)
+        draw.rectangle(
+            [(xmin + 1, ymin - th), (xmin + tw + 1, ymin)], fill=color)
+        draw.text((xmin + 1, ymin - th), text, fill=(255, 255, 255))
+    return im
diff --git a/requirements.txt b/requirements.txt
index c7d9d8093ad476856dcf643c08b10f8c5f252c30..2e0299ba1e2d4c9d434df4b4afa2974234dad4a5 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -5,4 +5,6 @@ tb-paddle
 tensorboard >= 1.15
 cython
 pycocotools
+opencv-python
+PyYAML
 shapely