Using a Pre-Trained Neural Network on MS CoCo DataSet using Mask R-CNN in TensorFlow / Keras.
I am always looking for use cases, always. Yesterday I gave a talk at the IoT Fusion Conference in Philadelphia about IoT. I had a Raspberry Pi taking web camera images. I mentioned I wasn't pointing at people because they may not want to be photographed. I am concerned with privacy. It gave me the idea it would be cool to block people or other things from images. Then on queue, a great library shows up in github. Thanks to Minimaxir's Person Blocker, I can now use Apache NiFi to pull remove people from images. I did a couple of minor tweaks to his code to add OpenCV image capture and output some JSON information on what happened. I am running this on an OSX laptop, but at some point I'll move it to a Raspberry Pi, TinkerBoard or NVidia Jetson TX1.
Please support this project: https://github.com/minimaxir/person-blocker
https://www.patreon.com/minimaxir
Using this pretrained neural network, we can block anything in this list of classes https://github.com/minimaxir/person-blocker/blob/master/classes.py.
You will need to download the MS Coco classes which aren't that large. (mask_rcnn_coco.h5)
This works without a GPU!!
To Install the Libraries:
pip install --upgrade pip pip install keras pip install tensorflow pip install opencv-python pip install uuid git clone https://github.com/minimaxir/person-blocker.git pip3 install -r requirements.txt
There are a bunch of requirements such as Python 3, recent TensorFlow (I used TF 1.7), Keras, Numpy, SkiImage, SCIPY, Pillow, Cython, H5PY, Matplotlib and imageio. I added uuid and json libraries. So you install them and get running.
The JSON produced as a record of the run has the following schema:
Schema
{ "type" : "record", "name" : "personblocker", "fields" : [ { "name" : "uuid", "type" : "string", "doc" : "Type inferred from '\"person_uuid_20180406203059f_b7ce1056-9d88-4e7f-b4dd-0e8c8d6e7086\"'" }, { "name" : "runtime", "type" : "string", "doc" : "Type inferred from '\"27\"'" }, { "name" : "host", "type" : "string", "doc" : "Type inferred from '\"server.local\"'" }, { "name" : "ts", "type" : "string", "doc" : "Type inferred from '\"2018-04-06 20:30:59\"'" }, { "name" : "ipaddress", "type" : "string", "doc" : "Type inferred from '\"10.1\"'" }, { "name" : "imagefilename", "type" : "string", "doc" : "Type inferred from '\"person_blocked_20180406203057\"'" }, { "name" : "originalfilename", "type" : "string", "doc" : "Type inferred from '\"images2/tx1_image_b9ebdd52-9a9f-45f0-b71c-a44c54f14b71_20180406203032.jpg\"'" } ] }
Example Output JSON
{"uuid": "person_uuid_20180406201647f_1d2c31bc-c232-4976-a350-747ffabf5afe", "runtime": "76", "host": "mymachine.local", "ts": "2018-04-06 20:16:47", "ipaddress": "10.1.1.12", "imagefilename": "person_blocked_20180406201632", "originalfilename": "images2/tx1_image_0309425f-12ca-4331-a810-21067cbaa8f2_20180406201531.jpg"}
run.sh
python3 -W ignore pb.py 2>/dev/null
Now You See Me (OpenCV Captured Image)
Now You Don't (Person Blocker - it did block a stove pipe)
There's also the option to produce a GIF that moves which is cool but takes time and space. I commented that out.
My modified example:
https://github.com/tspannhw/OpenSourceComputerVision
See run.sh and pb.py.
