Blurring Faces

Click the Open in Colab button to run the cookbook on Google Colab.

Introduction

In this cookbook we'll use a frame from a video of someone in a supermarket. We'll download this video via the supervision assets module. We'll then run inference on this frame using the hosted Roboflow API to fetch detections of faces utilizing an open source face detection model on Roboflow Universe. Finally, we'll use supervision to blur the detected faces.

Install packages

Let's quickly install the supervision package with the assets module, as well as the roboflow inference_sdk with pip. We'll also install tqdm to show a progress bar, but this is optional in production code.

python

!pip3 install -q supervision inference tqdm "Pillow<12"

Download Video and Extract Frame

In order to blur a face in a frame, we'll need a frame with a face in it. Let's download a video, and grab a frame in the middle of the video. I played around a little, and found that the 800th frame is great frame for us to test, since the customer is facing the camera. In this code, we're also using tqdm to display a progress bar of our script.

python

import supervision as sv

video = sv.download_assets(sv.VideoAssets.GROCERY_STORE)

# Seek directly to frame 800 using the start parameter (O(1) seek)
frame = next(sv.get_video_frames_generator(video, start=800))

sv.plot_image(frame)

Detecting Faces

Now that we've got our image we'll need a good face detecting model. For this task, there are already an impressive amount of open source models available on Roboflow Universe. After a little digging, this face detection model has over 1300 images. Some models, including this one, require a Roboflow API key. You can create a free account here. From there, you can find the key under Settings > Workspaces > Roboflow API. Let's give it a try.

python

import os
from inference_sdk import InferenceHTTPClient

try:
    from google.colab import userdata
    ROBOFLOW_API_KEY = userdata.get("ROBOFLOW_API_KEY") or ""
except ImportError:
    ROBOFLOW_API_KEY = os.environ.get("ROBOFLOW_API_KEY", "")

assert ROBOFLOW_API_KEY, "Set ROBOFLOW_API_KEY in Colab secrets or as env var"

client = InferenceHTTPClient(
    api_url="https://detect.roboflow.com",
    api_key=ROBOFLOW_API_KEY
)

results = client.infer(frame, model_id="face-detection-mik1i/18")

print(f"Detected {len(results['predictions'])} face(s)")
print(results)

Blurring the Face

Now that we're detecting faces, bluring them is easy with supervision. Let's pass our results into a Detections object and annotate the frame with a BlurAnnotator.

python

blur = sv.BlurAnnotator(kernel_size=100)

detections = sv.Detections.from_inference(results)

annotated_frame = blur.annotate(scene=frame.copy(), detections=detections)

sv.plot_image(annotated_frame)

Conclusion

With supervision, inference, and Roboflow Universe we were able to blur faces in minutes with an open source model. There are many other impressive use cases out there, so feel free to share in your own cookbooks. Happy building!