Implementation:PeterL1n BackgroundMattingV2 MattingBase

Knowledge Sources	BackgroundMattingV2 Real-Time High-Resolution Background Matting
Domains	Image_Matting, Computer_Vision, Deep_Learning
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for coarse alpha matte prediction from source-background image pairs provided by model/model.py.

Description

MattingBase is the first-stage matting model that extends the Base encoder-decoder with 6 input channels (concatenated source + background) and 37 output channels (1 alpha + 3 foreground + 1 error + 32 hidden). It supports ResNet50, ResNet101, and MobileNetV2 encoder backbones. The foreground is predicted as a residual added to the source image.

Usage

Use for coarse-resolution matting when refinement is not needed, or as the first stage when training the full pipeline. For inference, only the first two outputs (alpha, foreground) are typically used.

Code Reference

Source Location

Repository: BackgroundMattingV2
File: model/model.py
Lines: 61-98

Signature

class MattingBase(Base):
    """
    Coarse global matting at reduced resolution.

    Args:
        backbone: str - one of ['resnet50', 'resnet101', 'mobilenetv2']
    """
    def __init__(self, backbone: str):
        super().__init__(backbone, in_channels=6, out_channels=(1 + 3 + 1 + 32))

    def forward(
        self,
        src: Tensor,  # (B, 3, H, W) source image, RGB, 0-1
        bgr: Tensor   # (B, 3, H, W) background image, RGB, 0-1
    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
        """
        Returns:
            pha: (B, 1, H, W) alpha matte, 0-1
            fgr: (B, 3, H, W) foreground RGB, 0-1
            err: (B, 1, H, W) error map, 0-1
            hid: (B, 32, H, W) hidden features for refiner
        """

Import

from model import MattingBase

I/O Contract

Inputs

Name	Type	Required	Description
backbone	str	Yes	Encoder backbone: 'resnet50', 'resnet101', or 'mobilenetv2'
src	Tensor[B,3,H,W]	Yes	Source image (RGB, 0-1)
bgr	Tensor[B,3,H,W]	Yes	Background image (RGB, 0-1)

Outputs

Name	Type	Description
pha	Tensor[B,1,H,W]	Alpha matte prediction, clamped 0-1
fgr	Tensor[B,3,H,W]	Foreground prediction (src + residual), clamped 0-1
err	Tensor[B,1,H,W]	Error map prediction, clamped 0-1
hid	Tensor[B,32,H,W]	Hidden encoding for refinement stage (ReLU activated)

Usage Examples

Training

from model import MattingBase

model = MattingBase(backbone='resnet50')
pha, fgr, err, hid = model(src, bgr)  # Use all 4 outputs for training

Inference

model = MattingBase(backbone='resnet50')
model.load_state_dict(torch.load('checkpoint.pth'), strict=False)
model.eval()

with torch.no_grad():
    pha, fgr = model(src, bgr)[:2]  # Only need alpha and foreground

Related Pages

Implements Principle

Principle:PeterL1n_BackgroundMattingV2_Coarse_matting_prediction

Requires Environment

Environment:PeterL1n_BackgroundMattingV2_PyTorch_CUDA

Uses Heuristics

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment