In objectdash/settings.py set STATIC_ROOT and MEDIA_ROOT to point to the directories where static and user
upload content should be stored.:
STATIC_ROOT = '/home/baddad/workspace/objectdash/objectdash/web/static'
MEDIA_ROOT = '/home/baddad/workspace/objectdash/objectdash/web/static/media'
Install project requirements:
pipenv install .
Clone the tensorflow-models repository and add models/research to your python path to make the object_detection API
available.
git clone https://github.com/tensorflow/models
echo "export PYTHONPATH=`pwd`/models/research:$PYTHONPATH" >> ~/.bashrc
source ~/.bashrc
You'll need to compile the protobuf objects before the object_detection API can be used:
protoc ./object_detection/protos/*.proto --python_out = ./
Create the django database (a sqlite database in the project's root directory).
If you want to use another database see Django's documentation.
pipenv run ./manage.py migrate
Create an admin super user:
pipenv run ./manage.py createsuperuser
Run the development server:
$ pipenv run ./manage.py runserver
Use the webcam management command to detect in real-time against a camera.
Usage:
./manage.py webcam <path/to/frozen_inference_graph.pb> <path/to/labels.pbtxt> <num_classes>
Example:
./manage.py webcam ./data/models/faster_rcnn_resnet101_coco_2018_01_28/frozen_inference_graph.pb "./media/object_detectors/pb/ssd_resnet50_v1_fpn/labels.pbtxt" 90 --show_fps
Download an object detection model from the Tensorflow object detection model zoo and extract its contents. If you have your own object detection models, you can install that instead.
With the server running, visit http://localhost:8000/admin and log in using your admin account.
On the main admin page, under Object Detectors, click the Add link:
On the Add Object Detector page, specify the name, *.pb, and *.pbtxt file for your object detector, then click SAVE
You should now see your new model listed under Object Detectors on your server:
Visit http://localhost:8000/object_detection/single_image on your local instance to visit the Single Image submission page
Click Browse... to select which image you want annotated, set your confidence threshold and click Annotate.
The first time you submit an image the models marked active in the database are loaded into tensorflow. ]
This can cause a significant delay before results are displayed however any subsequent results should be significantly
faster (using faster-rcnn models, the first image takes 8-15 seconds on my local machine.
Subsequent requests take ~0.15 seconds to complete)
Once the results load you should see tabs for each active model, with the annotated image and extracted crops:
If you are using an object detector that provides masks (like mask-rcnn), these will be displayed instead of the bounding rect:
