Implementation:PeterL1n BackgroundMattingV2 VideoWriter

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

Overview

Concrete tool for writing PyTorch tensor batches to MP4 video files provided by inference_video.py.

Description

VideoWriter wraps OpenCV's cv2.VideoWriter to accept PyTorch float tensor batches and convert them to video frames. The conversion pipeline multiplies by 255, casts to uint8, permutes from (B,C,H,W) to (B,H,W,C), converts RGB to BGR, and writes each frame to an MP4 file using the mp4v codec.

Usage

Use when saving matting output as video files. One instance per output type (composite, alpha, foreground, error, refinement map).

Code Reference

Source Location

• Repository: BackgroundMattingV2
• File: inference_video.py
• Lines: 80-90

Signature

class VideoWriter:
    def __init__(
        self,
        path: str,       # Output MP4 file path
        frame_rate: float,  # Video frame rate
        width: int,       # Frame width
        height: int       # Frame height
    ):
        """Creates cv2.VideoWriter with mp4v codec."""

    def add_batch(self, frames: Tensor) -> None:
        """
        Write a batch of frames to the video.

        Args:
            frames: Tensor[B, 3, H, W] float32, range 0-1, RGB
        """

Import

# Defined in inference_video.py (not a separate module)
# Used internally within the video inference script

I/O Contract

Inputs

Name	Type	Required	Description
path	str	Yes	Output MP4 file path
frame_rate	float	Yes	Target video frame rate
width	int	Yes	Frame width in pixels
height	int	Yes	Frame height in pixels
frames	Tensor[B,3,H,W]	Yes	Batch of RGB float tensors (0-1) to write

Outputs

Name	Type	Description
.mp4 file	File	MP4 video file at specified path

Usage Examples

Writing Matting Output

# Create writers for each output type
pha_writer = VideoWriter('output/pha.mp4', vid.frame_rate, width, height)
com_writer = VideoWriter('output/com.mp4', vid.frame_rate, width, height)

# In inference loop
with torch.no_grad():
    for src, bgr in dataloader:
        pha, fgr = model(src.cuda(), bgr.cuda())[:2]
        com = fgr * pha + tgt_bgr * (1 - pha)

        pha_writer.add_batch(pha)
        com_writer.add_batch(com)

Related Pages

Implements Principle

• Principle:PeterL1n_BackgroundMattingV2_Video_output_writing