Skip to content

Latest commit

 

History

History
38 lines (26 loc) · 3.13 KB

design_considerations.md

File metadata and controls

38 lines (26 loc) · 3.13 KB

Detailed Design Choices

Web Framework

FastAPI was chosen for its support of asynchronous endpoints and WebSocket features, while being lightweight and requiring minimal setup. Compared to Django, FastAPI is better suited for this use case due to its simplicity and modern design.

Face Detection Model

There is a wide variety of face detection models available. At a high level, these can be categorized into: CPU-based models and GPU-based models.

Since portability is a key requirement, only CPU-based models were considered, as there is no guarantee that the server will have access to GPU acceleration.

Chosen Model: YuNet, a lightweight CPU-based model, was selected for this task. YuNet is efficient, offers reliable performance across diverse scenarios, and has low latency, making it ideal for near real-time face detection.
(Source: Medium article on face detection models)

Concurrency

The most performance-intensive endpoint is POST /image, which involves image processing, face detection, I/O operations, and WebSocket broadcasting.

  1. File Decoding and Validation:
    The input file is decoded and validated asynchronously within the FastAPI endpoint. If errors (e.g., invalid file format) are detected, they are returned to the client immediately.

  2. Background Tasks:
    Two background tasks are registered after successful validation:

    • Face Detection: This task uses YuNet to detect faces and saves the annotated image using OpenCV. It runs in a separate thread via FastAPI’s BackgroundTasks feature, ensuring that the main event loop remains unblocked.
      YuNet’s low latency allows it to perform face detection in near real-time, eliminating the need for parallel face detection processes.
      (Source: Springer article)
    • WebSocket Broadcasting: After face detection is complete, the URL of the generated image is broadcast asynchronously using the WebSocket connection.

  3. Immediate Response:
    The API sends a response to the client right after initiating the background tasks, without waiting for their completion. This approach ensures a responsive and non-blocking user experience.

Media Storage

Volatile disk storage within the Docker container was selected for storing processed images. This is a pragmatic and lightweight solution suitable for the scope of this task.

Model Initialization

To accurately detect faces in images of varying sizes, the FaceDetector class uses YuNet's setInputSize method to dynamically adjust the model’s input_size parameter based on the input image shape.

Concurrent requests to POST /image could lead to a race condition, as each request would overwrite the shared model configuration, potentially producing incorrect detection results.

The model is instantiated within the FaceDetector class, which is provided as a FastAPI dependency. This ensures that a new instance of the model is created for each request, isolating request processing and avoiding conflicts.