RESTful API for Amcrest cameras

USE conda install -- NOT pip install

conda install pip # might have an issue with conda

pip install flask
pip install opencv-python
conda install py-opencv
pip install pillow
pip install bs4
pip install lxml
pip install imutils
pip install matplotlib # req'd for the graph conversion pip install contextlib2
pip install jupyter
pip install boto3
pip install tf_slim
pip install PySocks
pip install requests
pip install hueber

Make sure you update ./bashrc
$LD_LIBRARY_PATH
/usr/local/cuda->cuda-10.0

Tensorflow or just the TFLite Interpreter? Well, you will be using a lot of tf.* utilities
pip install tensorflow-gpu==1.15

if 8100 - you have an old CPU - you need special TF build - get it off of jmduff/S3
tensorflow 1.15 required - or you'll get a graph error (incompatible versions)
python 3.6
pip install tensorflow-1.15.2-cp36-cp36m-linux_x86_64.whl

CUDA 10.0 & cuDNN 7.6.5.32 pip install tensorflow-gpu==1.15
CPU ~ 40%, GPU ~ 40%

CUDA 10.0 & cuDNN 7.6.5.32 conda activate security (tf 1.15)
CPU ~ 60%, GPU ~ 40%

intall kernel on Jupyter:
python -m ipykernel install --user --name=security

This will also compile the protobufs
bash ./install_tf_models.sh

you need the label map:
cp research/object_detection/data/mscoco_label_map.pbtxt ~/projects/camera-api/model/

you need a tflite model - easiest place to get that is from s3
you should have created it using the ssd-dag/UnderstandingTensorRT_ConvertGraph notebook

Which model type to use?

I started with tflite because it's simple. It does not use a GPU. TFLite is a stepping stone to EdgeTPU (Coral). Performance was about 0.1 sec per inference. Not bad but it adds up when you have 4+ image regions. the reolink2tflite.py program is highly deprecated and will no longer work - but, it's still there if you want to convert to tflite. This will show you how to use this model type.

This is a sensible way to go - but I did not. You convert the *.tflite model to an EdgeTPU model and then it will use the Coral. The Coral is fast - as fast as the GPU or faster. But, I did not use this. The interface to this model is different. Probably can figure this out by looking at other projects in the github account.

This was the preferred solution because I have GPUs. Requires TF 1.15
You may need to run:
python test_tf_inference.py
just to compile things - the security program may be slow starting the first time.

you may need to create the following subdirectories:

• faces/ images with faces, going to Rekognition
• snapshot/ camera images and inference.xml
• stream/cam0 - camN subdirectories if you stream

python security.py app_reolink12_tensorflow.json

using github labelimg
follow the install directions found in the README.md
$ conda activate labelimg
$ cd ~/projects/labelImg
$ python labelImg.py

hint: tar xvf tarball.tar.gz --strip-components=1

put annotations (xml) in annotations/
put images in jpg_images/

This was originally developed for tflite (mobilenet). Advantages:

• EdgeTPU compatible
• lightweight
• I knew what I was doing

Without using the Coral TPU, it was about 0.1 second execution (does not utilize GPU).

But, I have GPUs. Going to frozen graph - Advantages:

• easier to migrate to ResNet 50 (larger model) at some point
• GPUs with frozen graph are about 5x faster: 0.02 (probably same as a TPU stick in all fairness)
• probably slightly more accurate

While migrating to frozen graph - I migrated to multi-process (1 process/ camera). And things got messy and disorganized.

git clone https://github.com/Benehiko/ReolinkCameraAPI.git

pip install hueber git clone https://github.com/mbaltrusitis/hueber.git