English | 简体中文
These examples demonstrate how to use Intel® OpenVINOTM integration with Tensorflow to recognize and detect objects in images and videos.
- Python implementation for classification
- Python implementation for object detection
- C++ implementation for classification
- Classification demo uses Google's Inception v3 model to classify a given image, video, directory of images or camera input.
- Object detection demo uses YOLOv4 model converted from Darknet to detect objects in a given image, video, directory of images or camera input.
Before you proceed to run the examples, you will have to clone the openvino_tensorflow repository to your local machine. For this, run the following commands:
$ git clone https://github.com/openvinotoolkit/openvino_tensorflow.git
$ cd openvino_tensorflow
$ git submodule init
$ git submodule update --recursiveFor this example, we assume that you've already:
- Installed TensorFlow on your system
- Installed Intel® OpenVINOTM integration with Tensorflow on your system
Note: Refer to this page for python classification sample using TF1 Inception v3 model.
Refer to this page for quick installation using pip.
This demo uses TF Hub image classification model InceptionV3 and ImageNet labels for the possible classifications. In labels file, you'll find 1,000 categories that were used in the Imagenet competition.
Install the pre-requisites
$ cd <path-to-openvino_tensorflow-repository>/examples
$ pip3 install -r requirements.txtNow, you can run classification sample using the instructions below:
$ cd <path-to-openvino_tensorflow-repository>
$ python3 examples/classification_sample.pyclassification_sample.py does inference on the image and should output something similar to:
military uniform 0.79601693
mortarboard 0.02091024
academic gown 0.014557127
suit 0.009166162
comic book 0.007978318
Note: use --no_show flag to disable the application display window. By default the display window is enabled.
In this case, we are using the image of Admiral Grace Hopper. As you can see, the network correctly spots that she's wearing a military uniform, with a high score of 0.79.
Next, try it out by passing the path to your new input. You can provide either absolute or relative path to an image or video or directory of images. e.g.
$ python3 examples/classification_sample.py --input=<absolute-or-relative-path-to-your-input>If you add a new image or video (e.g, my_image.png or people-detection.mp4) to the existing <path-to-openvino_tensorflow-repository>/examples/data directory in the openvino_tensorflow repository, it will look like this:
$ python3 examples/classification_sample.py --input=examples/data/my_image.png
or
$ python3 examples/classification_sample.py --input=examples/data/people-detection.mp4For using camera as input use --input=0. Here '0' refers to the camera present at /dev/video0. If the camera is connected to a different port, change it appropriately.
To see more options for various backends (Intel® hardware), invoke:
$ python3 examples/classification_sample.py --helpFor this example, we assume that you've already:
- Installed TensorFlow on your system.
- Installed Intel® OpenVINOTM integration with Tensorflow on your system
Note: Refer to this page for the conversion of yolov4 darknet model and it's python object detection sample.
Refer to this page for quick installation using pip.
Install the pre-requisites
$ cd <path-to-openvino_tensorflow-repository>/examples
$ pip3 install -r requirements.txtThe TensorFlow Yolo v4 darknet model used in this demo is not packaged in the repository because of its size. So, follow the instructions below to convert the model from DarkNet to TensorFlow and download the labels and weights to the <path-to-openvino_tensorflow-repository>/examples/data directory in your cloned repo of openvino_tensorflow:
For Linux and macOS
$ cd <path-to-openvino_tensorflow-repository>/examples
$ chmod +x convert_yolov4.sh
$ ./convert_yolov4.shFor Windows
cd <path-to-openvino_tensorflow-repository>\examples
convert_yolov4.batOnce completed, the <path-to-openvino_tensorflow-repository>/examples/data folder will contain following files needed to run the object detection example:
- coco.names
- yolo_v4
Run the object detection example using the instructions below:
$ cd <path-to-openvino_tensorflow-repository>
$ python3 examples/object_detection_sample.pyNote:
- use
--no_showflag to disable the application display window. By default the display window is enabled. - use
--renameto rename the input image or the directory of the images based on the content of image, for example noexiftags-1person-1uniform-intelOpenVINO.jpg
This uses the default example image that comes with this repository, and should output something similar as below and the result is written to detections.jpg:
In this case, we're using the default image of Admiral Grace Hopper. As you can see, the network detects and draws the bounding box around the person correctly.
Next, try it out on your own image by passing the path to your new input. You can provide either absolute or relative path e.g.
$ python3 examples/object_detection_sample.py --input=<absolute-or-relative-path-to-your-image>If you add the new image (e.g, my_image.png) to the existing <path-to-openvino_tensorflow-repository>/examples/data directory in the openvino_tensorflow repository, it will look like this:
$ cd <path-to-openvino_tensorflow-repository>
$ python3 examples/object_detection_sample.py --input=examples/data/my_image.pngTo see more options for various backends (Intel® hardware), invoke:
$ python3 examples/object_detection_sample.py --helpNext, try it out on your own video file by passing the path to your input video. You can provide either absolute or relative path e.g.
$ python3 examples/object_detection_sample.py --input=<absolute-or-relative-path-to-your-video-file>If you add the new video (e.g, examples/data/people-detection.mp4) to the existing <path-to-openvino_tensorflow-repository>/examples/data directory in the openvino_tensorflow repository, it will look like this:
$ python3 examples/object_detection_sample.py --input=examples/data/people-detection.mp4For using camera as input use --input=0. Here '0' refers to the camera present at /dev/video0. If the camera is connected to a different port, change it appropriately.
Note: The results with input as an image or a directory of images, are written to output images. For video or camera input, use the application display window for the results.
For running C++ examples, we need to build TensorFlow framework from source since examples have a dependency on the TensorFlow libraries.
Before you start building from source, you have to make sure that the prerequisites are installed.
The TensorFlow model used in this demo is not packaged in the repo because of its size. So, download the model to the <path-to-openvino_tensorflow-repository>/examples/data directory in your cloned repo of openvino_tensorflow and extract the file:
- Download model
$ curl -L "https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz" |
tar -C <path-to-openvino_tensorflow-repository>/examples/data -xz- Run the following commands to build openvino_tensorflow with samples:
$ cd <path-to-openvino_tensorflow-repository>
$ python3 build_tf.py --output_dir <path-to-tensorflow-dir>
$ python3 build_ovtf.py --use_tensorflow_from_location <path-to-tensorflow-dir>For detailed build instructions read BUILD.md.
- Now, a binary executable for classification_sample is built. Update the LD_LIBRARY_PATH and run the sample:
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<path-to-openvino_tensorflow-repository>/build_cmake/artifacts/lib:<path-to-openvino_tensorflow-repository>/build_cmake/artifacts/tensorflow
$ ./build_cmake/examples/classification_sample/infer_image-
Download following model "https://storage.googleapis.com/download.tensorflow.org/models/inception_v3_2016_08_28_frozen.pb.tar.gz" uncompress it and copy it to <path-to-openvino_tensorflow-repository>\examples\data
-
Run the following commands to build openvino_tensorflow with samples
cd <path-to-openvino_tensorflow-repository>
python build_ovtf.py --use_openvino_from_location="C:\Program Files (x86)\Intel\openvino_2022.3.0" --use_tensorflow_from_location="\path\to\directory\containing\tensorflow\"- Now, a binary executable for classification_sample is built. Update the PATH and run the sample:
set PATH=%PATH%;<path-to-openvino_tensorflow-repository>\build_cmake\artifacts\lib;<path-to-openvino_tensorflow-repository>\build_cmake\artifacts\tensorflow
build_cmake\examples\classification_sample\Release\infer_image.exeThis uses the default example image that comes with this repository, and should output something similar to:
military uniform (653): 0.834306
mortarboard (668): 0.0218693
academic gown (401): 0.010358
pickelhaube (716): 0.00800814
bulletproof vest (466): 0.00535091
In this case, we're using the default image of Admiral Grace Hopper. As you can see the network correctly spots she's wearing a military uniform, with a high score of 0.8.
Next, try it out on your own image by passing the path to your new input. You can provide either absolute or relative path e.g.
$ ./build_cmake/examples/classification_sample/infer_image --image=<absolute-or-relative-path-to-your-image>If you add the new image (e.g, my_image.png) to the existing <path-to-openvino_tensorflow-repository>/examples/data directory in the openvino_tensorflow repository, it will look like this:
$ ./build_cmake/examples/classification_sample/infer_image --image=examples/data/my_image.pngTo see more options for various backends (Intel® hardware), invoke:
$ ./build_cmake/examples/classification_sample/infer_image --helpNote: In the above samples a warm-up run is executed first and then inference time is measured on the subsequent runs. The execution time of first run is in general higher compared to the next runs as it includes many one-time graph transformations and optimizations steps.