The goal of Focused Vision is to capture an object of interest in as much detail as possible and to do all necessary steps on-device. It excels when the object occupies only a small part of the image - whether because it is physically small, relatively far from the camera, or both.
This application compares a naive face-detection approach with two Focused Vision approaches using person detection and tiling.
Note: This example works only on RVC4 in standalone mode.
Detects faces on downscaled (320x240px) low-res RGB
This approach uses a two-stage neural-network pipeline for significantly improved face-detection reliability.
-
Stage 1 – Person Detection: A person is detected on the full 2000×2000 high-resolution image. The person region is then cropped at full detail.
-
Stage 2 – Face Detection on the Crop: The face detector runs on the cropped person image, downscaled to 320×240—the same input size used in the naïve approach.
The key difference is where the downscaling happens.
• In the naïve approach, the entire image is downscaled to 320×240 before face detection, causing the face to become small and lose detail.
• In the NN chaining approach, only the high-resolution person crop is downscaled, preserving far more facial detail in the model’s input.
As a result, the face detector receives a clearer, more information-rich image, which substantially increases the chance of successful face detection.
Tiling is a brute-force approach that increases face-detection reliability by splitting the high-resolution image into multiple overlapping regions (tiles). Each tile is then independently downscaled to 320×240 and passed through the face-detection model.
Because each tile represents only a small portion of the original image, more facial detail is preserved compared to downscaling the entire frame at once—making detection more likely.
This method is particularly useful when the NN Model Chaining technique cannot be applied. However, it is computationally expensive. In this application, a 2×2 tiling pattern is used, producing 4 tiles. Since the face-detection model must run on every tile, a 15 FPS camera effectively requires the neural network to process 4× the number of frames, greatly increasing processing load.
Running this example requires a Luxonis device connected to your computer. Refer to the documentation to set up your device if you haven't already.
To run the example in this mode, first install the oakctl tool using the installation instructions here.
The app can then be run with:
oakctl connect <DEVICE_IP>
oakctl app run .Once the app is built and running you can access the DepthAI Viewer locally by opening https://<OAK4_IP>:9000/ in your browser (the exact URL will be shown in the terminal output).
- You can upload the oakapp to Luxonis Hub via
oakctl. - Then you can remotely open the App UI via the App detail page (WebRTC).
This example runs with default argument values. If you’d like to customize it, you can use it as a starting point and modify the source code, which is available here. For more examples, check out our app store or browse the oak-examples repository.