English | 中文

Official repository for the paper Robust High-Resolution Video Matting with Temporal Guidance. RVM is specifically designed for robust human video matting. Unlike existing neural models that process frames as independent images, RVM uses a recurrent neural network to process videos with temporal memory. RVM can perform matting in real-time on any videos without additional inputs. It achieves 4K 76FPS and HD 104FPS on an Nvidia GTX 1080 Ti GPU. The project was developed at ByteDance Inc.

• [Nov 03 2021] Fixed a bug in train.py.
• [Sep 16 2021] Code is re-released under GPL-3.0 license.
• [Aug 25 2021] Source code and pretrained models are published.
• [Jul 27 2021] Paper is accepted by WACV 2022.

Watch the showreel video (YouTube, Bilibili) to see the model's performance.

All footage in the video are available in Google Drive.

• Webcam Demo: Run the model live in your browser. Visualize recurrent states.
• Colab Demo: Test our model on your own videos with free GPU.

We recommend MobileNetv3 models for most use cases. ResNet50 models are the larger variant with small performance improvements. Our model is available on various inference frameworks. See inference documentation for more instructions.

All models are available in Google Drive and Baidu Pan (code: gym7).

1. Install dependencies:

pip install -r requirements_inference.txt

2. Load the model:

import torch
from model import MattingNetwork

model = MattingNetwork('mobilenetv3').eval().cuda()  # or "resnet50"
model.load_state_dict(torch.load('rvm_mobilenetv3.pth'))

3. To convert videos, we provide a simple conversion API:

from inference import convert_video

convert_video(
    model,                           # The model, can be on any device (cpu or cuda).
    input_source='input.mp4',        # A video file or an image sequence directory.
    output_type='video',             # Choose "video" or "png_sequence"
    output_composition='com.mp4',    # File path if video; directory path if png sequence.
    output_alpha="pha.mp4",          # [Optional] Output the raw alpha prediction.
    output_foreground="fgr.mp4",     # [Optional] Output the raw foreground prediction.
    output_video_mbps=4,             # Output video mbps. Not needed for png sequence.
    downsample_ratio=None,           # A hyperparameter to adjust or use None for auto.
    seq_chunk=12,                    # Process n frames at once for better parallelism.
)

4. Or write your own inference code:

from torch.utils.data import DataLoader
from torchvision.transforms import ToTensor
from inference_utils import VideoReader, VideoWriter

reader = VideoReader('input.mp4', transform=ToTensor())
writer = VideoWriter('output.mp4', frame_rate=30)

bgr = torch.tensor([.47, 1, .6]).view(3, 1, 1).cuda()  # Green background.
rec = [None] * 4                                       # Initial recurrent states.
downsample_ratio = 0.25                                # Adjust based on your video.

with torch.no_grad():
    for src in DataLoader(reader):                     # RGB tensor normalized to 0 ~ 1.
        fgr, pha, *rec = model(src.cuda(), *rec, downsample_ratio)  # Cycle the recurrent states.
        com = fgr * pha + bgr * (1 - pha)              # Composite to green background. 
        writer.write(com)                              # Write frame.

5. The models and converter API are also available through TorchHub.

# Load the model.
model = torch.hub.load("PeterL1n/RobustVideoMatting", "mobilenetv3") # or "resnet50"

# Converter API.
convert_video = torch.hub.load("PeterL1n/RobustVideoMatting", "converter")

Please see inference documentation for details on downsample_ratio hyperparameter, more converter arguments, and more advanced usage.

Please refer to the training documentation to train and evaluate your own model.

Speed is measured with inference_speed_test.py for reference.

• Note 1: HD uses downsample_ratio=0.25, 4K uses downsample_ratio=0.125. All tests use batch size 1 and frame chunk 1.
• Note 2: GPUs before Turing architecture does not support FP16 inference, so GTX 1080 Ti uses FP32.
• Note 3: We only measure tensor throughput. The provided video conversion script in this repo is expected to be much slower, because it does not utilize hardware video encoding/decoding and does not have the tensor transfer done on parallel threads. If you are interested in implementing hardware video encoding/decoding in Python, please refer to PyNvCodec.

• Shanchuan Lin
• Linjie Yang
• Imran Saleemi
• Soumyadip Sengupta

• NCNN C++ Android (@FeiGeChuanShu)
• lite.ai.toolkit (@DefTruth)
• Gradio Web Demo (@AK391)
• Unity Engine demo with NatML (@natsuite)
• MNN C++ Demo (@DefTruth)
• TNN C++ Demo (@DefTruth)