BlurScene

BlurScene is a model to anonymize faces and license plates of traffic data.

Table of Contents

1. Setup
   1. Environment Setup
2. Inference
   1. Configuration
   2. Inference Script
   3. Server
   4. Docker Container

Setup

Environment Setup

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

should suffice given a Python version >= 3.10. The code was written using Python 3.10.12.

Inference

Configuration

The configuration for inference, i.e. for the files inference.py and run_server.py, can be found in config/inference.yaml.

The paths to the model weights (.pt) and model configuration (.yaml, a hydra config) have to be given separately. Since the model configuration file completely defines the model architecture, one has to be sure that the conf fits the weights. Additionally, .dockerignore whitelists a weights directory, so it is easiest to put model weights and config there. The model device can be specified separately in config/inference.yaml.

Additional processing steps can be configured in the processing section. It is possible to send the image in multiple variants through the model and gather the detections from each variant. There a two options:

• mirror_image which mirrors the image horizontally,
• enlarged_regions_n which slices the image into $n^2$ parts (i.e. the _n defines the slices per dimension).

Furthermore, there are the usual post-processing options, i.e.

• pre_merge_score_threshold, post_merge_score_threshold throw away predictions below the given scores
• merging_method sets a merging algorithm, one of
  • Weighted Box Fusion, which computes an average box from overlapping boxes,
  • Non-Maximum Suppression, which only keeps the highest scoring of overlapping boxes, and
  • Non-Maximum Merging, which produces the enveloping box.

The merge_iou_threshold defines at which overlap (Intersection over Union, IoU) boxes must be merged. area_method changes how the box area for the IoU is computed, either

• int is $A = (x_1 - x_0 + 1) * (y_1 - y_0 + 1)$
• float is $A = (x_1 - x_0) * (y_1 - y_0)$.

At last one can define the logging level (the usual logging-module levels) and the log format.

Inference Script

inference.py is a module used by run_server.py. However, for testing purposes it is executable. Simply running python inference.py path/to/image.jpg should load the model, warm up/compile it, and predict bounding boxes for the given image.

Server

The run_server.py script starts a local Flask server for testing. For proper worker handling, etc. the server should be started using gunicorn, gunicorn --config=config/gunicorn.py run_server:app. gunicorn can be configured in config/gunicorn.py (see the gunicorn documentation for details). Here it is pre-configured to query the environment variables WORKERS, WORKER_TIMEOUT, PORT and LOG_LEVEL. By default the gunicorn server has the address 0.0.0.0:5000, i.e. it opens the port 5000 to the network side. Flask on the other hand runs with its default configuration and opens port 5000 only for connections from localhost.

The anonymization endpoint is /anonymize and setup can be tested with e.g. curl -H "Content-Type: image/jpeg" --data-binary @test.jpg http://localhost:5000/anonymize --output returned_image.jpg.

Docker Container

The docker container is built by docker build . -t $image_name and can be run with e.g. docker run -d --gpus all -p 5000:5000 --name $container_name $image_name. With docker logs -f $container_name one can monitor the logs and see when the model is ready.