tf-object-detector

A small component for you to start processing images and recognizing their content. Based on the amazing work by https://github.com/webrtcHacks/tfObjWebrtc

Disclaimer

This is our effort to dockerize the work done by https://github.com/webrtcHacks/tfObjWebrtc maintaining most of the original code. So why didn't we use his server image? (or better yet: why did we create this image in the first place?):

• We included waitress to mediate requests to the flask API;
• We forced all working versions of python packages, base images and so on, to keep this as stable as possible;
• We turned this into a self-contained (micro)service so that you can use it in your architecture without any hussle;
• You can even modify or add things to the base docker image, to tune it further if you would like;
• Open source the full code, Dockerfile included;
• Removed sample code from within the flask api;

It was our sole intention to contribute to the community what we had to do to get this to work on our own projects ❤️

As you can see here, object detection runs in realtime returning a JSON of the objects detected in each frame. You simply have to post an image to the service and it will return something like this:

[
    {
        "class_name": "car",
        "height": 0.5318315029144287,
        "score": 0.7755758762359619,
        "width": 0.7912412881851196,
        "x": 0.5782610177993774,
        "y": 0.3312181532382965
    },
    {
        "class_name": "car",
        "height": 0.4864712953567505,
        "score": 0.512231171131134,
        "width": 0.5383723974227905,
        "x": 0.4695549011230469,
        "y": 0.4001152515411377
    },
    {
        "class_name": "car",
        "height": 0.4555974006652832,
        "score": 0.4407155513763428,
        "width": 0.8605666756629944,
        "x": 0.7803468108177185,
        "y": 0.34807342290878296
    },
    {
        "class_name": "person",
        "height": 1.0,
        "score": 0.43984636664390564,
        "width": 0.5317587852478027,
        "x": 0.019111961126327515,
        "y": 0.0471099317073822
    }
]

The class_name property refers to the type of object detected within the chosen dataset (by default, we're using COCO dataset)

Table of Contents

• Quickstart
• Prerequisites
• Contributing
• Contributors
• Support
• License

---

TLDR; Quickstart

If you want, you can use this in a single step using the DockerHub registry:

docker run -d -p 8080:8080 --name tf-object-detector triedeti/tf-object-detector

---

Quickstart

• Clone or download this repo onto your machine;
• Open up a command line, and build the image:

docker build . --tag tf-object-detector

• Have some patience, this may take a while...
• You should see something like this on your machine:

Successfully built 
Successfully tagged tf-object-detector:latest

• Run a new container with the freshly built image:

docker run -d -p 8080:8080 --name tf-object-detector tf-object-detector

Hint: if it doesn't work, check if you have the 8080 port available

• If you open up your browser in http://localhost:8080 you will see something like this:

{
   "about": "tf-object-detector v1.0.0",
   "now": "2020-03-30 10:00:00.00000",
   "tensorflow_model": "ssd_mobilenet_v1_coco_2017_11_17",
   "tensorflow_version": "1.15.2"
}

• Now, you can simply open up the docs/video.html file on your preferred browser and click play on the video. Your newly created service will start processing each frame on the video!

---

Prerequisites

For this image to work, you only need Docker installed on your machine or server. That's all folks!

---

Contributing

Any contributions to this project are more than welcome. Feel free to reach us and we will gladly include any improvements or ideas that you may have. Please, fork this repository, make any changes and submit a Pull Request and we will get in touch!

Contributors

Jorge Santos	Leonel Dias	Licínio Carvalho
github.com/jdsantos	github.com/leoneljdias	https://github.com/LCNCC

Support

The easiest way to seek support is by submiting an issue on this repo. Also, reach out to us at one of the following places!

• Website at TriedeTI
• Twitter at @triedeti
• Facebook at @triedeti

---

License

• MIT license