This repository provides functionality to inpaint specific objects in video using a mmdetection.
This example is to inpaint human.
| before | after |
|---|---|
 |
 |
By the way, my amazing wife is in the video
This repository requires the following tools:
- windows
- CUDA_11.8
- python3.10.X
- ffmpeg
- Miniconda
It may work with other versions, but it is not guaranteed.
This repository uses ffmpeg command.
Please install from the official FFmpeg website.
Execute following command and confirm CUDA version is 11.8 or higher.
$ nvidia-smi
Execute following command and confirm CUDA version is 11.8 or higher.
Verify that CUDA_11.8 is built.
$ nvcc -V
Download Miniconda in official site https://docs.anaconda.com/miniconda/
Refer MMDetection official site. https://mmdetection.readthedocs.io/en/latest/get_started.html
Note
This document is for OS X. Pytorch install command is a bit different from the command mentioned in this document. Quote from Pytorch official site. https://pytorch.org/get-started/previous-versions/
Note
As of July 23, 2024, MMDetection appeared to does not supported mmcv >= 2.2.0.
So, mmcv version you should install is 2.1.0, and mmcv 2.1.0 is only supported 1.8.X <= torch <= 2.1.X.
To install MMDetection, it is better to install torch anb mmcv using the following commands.
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda={your cuda version} -c pytorch -c nvidia
pip install -U openmim
mim install "mmcv==2.1.0"
Following the document, download R-50-FPN pytorch 2x model. https://github.com/open-mmlab/mmdetection/tree/main/configs/mask_rcnn
Execute following command.
$ python main.py configs/mask_rcnn/mask-rcnn_r50_fpn_2x_coco.py mask_rcnn_r50_fpn_2x_coco_bbox_mAP-0.392__segm_mAP-0.354_20200505_003907-3e542a40.pth {your video data}