Frigate Classifier

Identify breeds/species of dogs and birds detected by blakeblackshear/frigate. This will try to identify a bird or dog and add a Frigate sublabel to that event.

Setup

Create a config.yml file in your docker volume with the following contents:

frigate:
  frigate_url: http://127.0.0.1:5000
  mqtt_server: 127.0.0.1
  mqtt_auth: false
  mqtt_username: username
  mqtt_password: password
  main_topic: frigate
  sublabel_score: true
  camera:
    - birdcam
bird_classification:
  threshold: 0.7
dog_classification:
  threshold: 0.7
logger_level: INFO

Update your frigate url, mqtt server settings. If you are using mqtt authentication, update the username and password. Update the camera name(s) to match the camera name in your frigate config.

You can also update the threshold for the bird and dog classification. The threshold is the minimum confidence level for the classification to be considered valid. The default is 0.7.

If you dont want to classify birds or dogs, you can remove the bird_classification or dog_classification sections from the config.

Running

docker run -v /path/to/config:/config -e TZ=America/New_York -it --rm --name frigate_classifier lmerza/frigate_classifier:latest

or using docker-compose:

services:
  frigate_classifier:
    image: lmerza/frigate_classifier:latest
    container_name: frigate_classifier
    volumes:
      - /path/to/config:/config
    restart: unless-stopped
    environment:
      - TZ=America/New_York

https://hub.docker.com/r/lmerza/frigateclassifier

Debugging

set logger_level in your config to DEBUG to see more logging information:

logger_level: DEBUG

Logs will be in /config/frigateclassifier.log

Training Dog Dataset

conda env create -f environment.yml
conda activate fc_env
pip install -r requirements.txt

curl -O http://vision.stanford.edu/aditya86/ImageNetDogs/images.tar
mkdir -p dog_images
tar -xf images.tar -C dog_images --strip-components=1

python format_dog_dataset.py
python train_dog_model.py

or

nohup python3 train_dog_model.py > output.log 2>&1 &

Attributions

The dog model was trained by Stanford Dogs Dataset