Implementation:PeterL1n BackgroundMattingV2 Camera

Knowledge Sources	BackgroundMattingV2
Domains	Video_Processing, Realtime_Systems
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for threaded webcam frame capture provided by inference_webcam.py.

Description

Camera wraps cv2.VideoCapture with a background reading thread. The thread continuously reads frames and stores the latest one behind a threading.Lock. The main thread calls read() to get the most recent frame without blocking on camera I/O. The camera's resolution is configured via OpenCV properties, and the actual negotiated resolution is stored in width and height attributes.

Usage

Use in the real-time webcam matting demo. Instantiate once at startup with desired resolution, then call read() in the matting loop.

Code Reference

Source Location

• Repository: BackgroundMattingV2
• File: inference_webcam.py
• Lines: 57-83

Signature

class Camera:
    def __init__(
        self,
        device_id: int = 0,
        width: int = 1280,
        height: int = 720
    ):
        """
        Args:
            device_id: Camera device index (default 0)
            width: Requested frame width
            height: Requested frame height

        Attributes:
            width: int - Actual negotiated frame width
            height: int - Actual negotiated frame height
        """

    def read(self) -> numpy.ndarray:
        """Returns latest frame as BGR numpy array (thread-safe copy)."""

    def __exit__(self, exec_type, exc_value, traceback) -> None:
        """Releases the camera capture."""

Import

# Defined in inference_webcam.py (not a separate module)

I/O Contract

Inputs

Name	Type	Required	Description
device_id	int	No	Camera device index (default 0)
width	int	No	Requested frame width (default 1280)
height	int	No	Requested frame height (default 720)

Outputs

Name	Type	Description
read()	numpy.ndarray	Latest camera frame as BGR uint8 array (H, W, 3)
.width	int	Actual camera width
.height	int	Actual camera height

Usage Examples

Webcam Matting Loop

cam = Camera(device_id=0, width=1280, height=720)
print(f"Camera: {cam.width}x{cam.height}")

# Background capture
frame = cam.read()
bgr = cv2_frame_to_cuda(frame)  # Convert BGR numpy to CUDA tensor

# Matting loop
while True:
    frame = cam.read()
    src = cv2_frame_to_cuda(frame)
    pha, fgr = model(src, bgr)[:2]
    # ... display result ...

Related Pages

Implements Principle

• Principle:PeterL1n_BackgroundMattingV2_Webcam_capture