Horizontal Text Detection in Real-Time¶

This demo presents a use case with a client written in python which captures camera frames and performs text spotting analysis via gRPC requests to OVMS. The client visualizes the results as a boxes depicted on the original image frames using OpenCV in real-time. The client can work efficiently also over slow internet connection with long latency thanks to image data compression and parallel execution for multiple frames.

horizontal text detection

Download horizontal text detection model from OpenVINO Model Zoo¶

curl -L --create-dir https://storage.openvinotoolkit.org/repositories/open_model_zoo/2022.1/models_bin/2/horizontal-text-detection-0001/FP32/horizontal-text-detection-0001.bin -o horizontal-text-detection-0001/1/horizontal-text-detection-0001.bin https://storage.openvinotoolkit.org/repositories/open_model_zoo/2022.1/models_bin/2/horizontal-text-detection-0001/FP32/horizontal-text-detection-0001.xml -o horizontal-text-detection-0001/1/horizontal-text-detection-0001.xml

tree horizontal-text-detection-0001
horizontal-text-detection-0001
└── 1
    ├── horizontal-text-detection-0001.bin
    └── horizontal-text-detection-0001.xml

Start the OVMS container:¶

docker run -d -u $(id -u):$(id -g) -v $(pwd)/horizontal-text-detection-0001:/model -p 9000:9000 openvino/model_server:latest \
--model_path /model --model_name text --port 9000 --layout NHWC:NCHW

Run the client¶

Clone the repository and enter horizontal_text_detection directory

git clone https://github.com/openvinotoolkit/model_server.git
cd model_server/demos/horizontal_text_detection/python

Install required packages:

pip3 install -r requirements.txt

Start the client

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000

You can also change the camera ID:

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000 --video_source 0

Or choose to work with video file as well:

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000 --video_source ~/video.mp4

Example output:

Initializing requesting thread index: 0
Initializing requesting thread index: 1
Initializing requesting thread index: 2
Initializing requesting thread index: 3
Launching requesting thread index: 0
Launching requesting thread index: 1
Launching requesting thread index: 2
Launching requesting thread index: 3
ThreadID:   0; Current FPS:    31.25; Average FPS:    25.64; Average latency:   140.98ms
ThreadID:   1; Current FPS:    31.23; Average FPS:    25.67; Average latency:   136.36ms
ThreadID:   2; Current FPS:    29.41; Average FPS:    25.70; Average latency:   130.88ms
ThreadID:   3; Current FPS:    30.30; Average FPS:    25.73; Average latency:   135.65ms
...

NOTE: Video source is cropped to 704x704 resolution to match model input size.

Recognize Detected Text with OCR Pipeline¶

Optical Character Recognition (OCR) pipeline based on horizontal text detection model, text recognition combined with a custom node implementation can be used with the same python script used before. OCR pipeline provides location of detected text boxes on the image and additionally recognized text for each box.

horizontal text detection using OCR pipeline

Prepare workspace to run the demo¶

To successfully deploy OCR pipeline you need to have a workspace that contains:

horizontal text detection and text recognition models
Custom node for image processing
Configuration file

Clone the repository and enter horizontal_text_detection directory

git clone https://github.com/openvinotoolkit/model_server.git
cd model_server/demos/horizontal_text_detection/python

You can prepare the workspace that contains all the above by just running make command. Since custom node used in this demo is included in OpenVINO Model Server image you can either use the custom node from the image, or build one.

If you just want to quickly run this demo and use already compiled custom node, run:

make

Directory structure (without custom node)¶

Once the make procedure is finished, you should have workspace directory ready with the following content.

workspace/
├── config.json
├── horizontal-text-detection-0001
│   └── 1
│       ├── horizontal-text-detection-0001.bin
│       └── horizontal-text-detection-0001.xml
└── text-recognition-0014
    └── 1
        ├── text-recognition-0014.bin
        └── text-recognition-0014.xml

If you modified the custom node or for some other reason, you want to have it compiled and then attached to the container, run:

make BUILD_CUSTOM_NODE=true BASE_OS=ubuntu

Directory structure (with custom node)¶

Once the make procedure is finished, you should have workspace directory ready with the following content.

workspace/
├── config.json
├── horizontal-text-detection-0001
│   └── 1
│       ├── horizontal-text-detection-0001.bin
│       └── horizontal-text-detection-0001.xml
├── lib
│   └── libcustom_node_horizontal_ocr.so
└── text-recognition-0014
    └── 1
        ├── text-recognition-0014.bin
        └── text-recognition-0014.xml

Deploying OVMS¶

Deploy OVMS with faces analysis pipeline using the following command:

docker run -p 9000:9000 -d -v ${PWD}/workspace:/workspace openvino/model_server --config_path /workspace/config.json --port 9000

Sending Request to the Server¶

Install python dependencies:

pip3 install -r requirements.txt

Start the client

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000 --use_case ocr

You can also change the camera ID:

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000 --use_case ocr --video_source 0

Or choose to work with video file as well:

python3 horizontal_text_detection.py --grpc_address localhost --grpc_port 9000 --use_case ocr --video_source ~/video.mp4

Example output:

Initializing requesting thread index: 0
Initializing requesting thread index: 1
Initializing requesting thread index: 2
Initializing requesting thread index: 3
Launching requesting thread index: 0
Launching requesting thread index: 1
Launching requesting thread index: 2
Launching requesting thread index: 3
ThreadID:   0; Current FPS:    31.25; Average FPS:    25.64; Average latency:   140.98ms
ThreadID:   1; Current FPS:    31.23; Average FPS:    25.67; Average latency:   136.36ms
ThreadID:   2; Current FPS:    29.41; Average FPS:    25.70; Average latency:   130.88ms
ThreadID:   3; Current FPS:    30.30; Average FPS:    25.73; Average latency:   135.65ms
...

RTSP Client¶

Build docker image containing rtsp client along with its dependencies The rtsp client app needs to have access to RTSP stream to read from and write to.

Example rtsp server mediamtx

docker run --rm -d -p 8080:8554 -e RTSP_PROTOCOLS=tcp bluenviron/mediamtx:latest

Then write to the server using ffmpeg, example using video or camera

ffmpeg -stream_loop -1 -i ./video.mp4 -f rtsp -rtsp_transport tcp rtsp://localhost:8080/channel1

ffmpeg -f dshow -i video="HP HD Camera" -f rtsp -rtsp_transport tcp rtsp://localhost:8080/channel1

Build the docker image with the python client for video stream reading an remote analysis:

docker build ../../common/stream_client/ -t rtsp_client

Start the client¶

Command

docker run -v $(pwd):/workspace rtsp_client --help
usage: rtsp_client.py [-h] [--grpc_address GRPC_ADDRESS]
                      [--input_stream INPUT_STREAM]
                      [--output_stream OUTPUT_STREAM]
                      [--model_name MODEL_NAME] [--width WIDTH]
                      [--height HEIGHT] [--input_name INPUT_NAME] [--verbose]
                      [--benchmark]
                      [--limit_stream_duration LIMIT_STREAM_DURATION]
                      [--limit_frames LIMIT_FRAMES]

options:
  -h, --help            show this help message and exit
  --grpc_address GRPC_ADDRESS
                        Specify url to grpc service
  --input_stream INPUT_STREAM
                        Url of input rtsp stream
  --output_stream OUTPUT_STREAM
                        Url of output rtsp stream
  --model_name MODEL_NAME
                        Name of the model
  --width WIDTH         Width of model's input image
  --height HEIGHT       Height of model's input image
  --input_name INPUT_NAME
                        Name of the model's input
  --verbose             Should client dump debug information
  --benchmark           Should client collect processing times
  --limit_stream_duration LIMIT_STREAM_DURATION
                        Limit how long client should run
  --limit_frames LIMIT_FRAMES
                        Limit how many frames should be processed

Usage example

docker run --network="host" -v $(pwd):/workspace rtsp_client --grpc_address localhost:9000 --input_stream 'rtsp://localhost:8080/channel1' --output_stream 'rtsp://localhost:8080/channel2'

Then read rtsp stream using ffplay

ffplay -pixel_format yuv420p -video_size 704x704 -rtsp_transport tcp rtsp://localhost:8080/channel2

One might as well use prerecorded video and schedule it for inference. Replace horizontal_text.mp4 with your video file.

docker run --network="host" -v $(pwd):/workspace rtsp_client --grpc_address localhost:9000 --input_stream 'workspace/horizontal_text.mp4' --output_stream 'workspace/output.mp4'