Awesome

InterpAny-Clearer

:rocket: <u style="color: hotpink; text-decoration: underline dotted hotpink;">[ECCV2024 Oral] Clearer Frames,</u> <u style="color: dodgerblue; text-decoration: underline dotted dodgerblue;">Anytime</u>: Resolving Velocity Ambiguity in Video Frame Interpolation

by Zhihang Zhong^1,*, Gurunandan Krishnan², Xiao Sun¹, Yu Qiao¹, Sizhuo Ma^2,†, and Jian Wang^2,†

^*First author, ^†Co-corresponding authors

¹Shanghai AI Laboratory, OpenGVLab, ²Snap Inc.

:point_right: project page
:point_right: interactive demo
:point_right: OpenXLab demo
:point_right: arXiv
:point_right: slides

Please leave a 🌟 if you like this project! 🔥🔥🔥

News

• :tada: 2024-08-12: Luckily, this work is recognized as Oral by ECCV2024! 🏁
• :tada: 2024-07-01: This work is accepted to ECCV2024! 🎆
• :tada: 2024-05-02: Our technology is used by CCTV5 and CCTV5+ for slow motion demonstrations of athletes jumping in the 2024 Thomas & Uber Cup! 🔥
• :tada: 2023-11-28: We have added an interface for video inference to the interactive demo, and uploaded checkpoints trained with the LPIPS loss.

Application in CCTV

https://github.com/zzh-tech/InterpAny-Clearer/assets/68437458/a4b0bc95-d051-45ac-aaf1-400266a290d2

TL;DR:

We addressed velocity ambiguity in video frame interpolation through innovative distance indexing and iterative reference-based estimation strategies, resulting in:
<b style="color: orangered">Clearer anytime frame interpolation</b> & <b style="color: orangered">Manipulated interpolation of anything</b>

Time indexing vs. Distance indexing

Comparison of x128 interpolation using only 2 frames as inputs:

[D]: distance indexing
[R]: iterative reference-based estimation

The results of [D,R] RIFE-vgg are perceptually clearer, but may suffer from undesirable distortions (see second row). We recommend using [D,R] RIFE for more stable results.

Preparation

Environment installation:

You can try Anaconda or Docker to setup the environment.

Anaconda

conda create -n InterpAny python=3.8
conda activate InterpAny
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install -r requirements.txt

Docker

You can build a docker image with all dependencies installed.
See docker/README.md for more details.

Download checkpoints

We provide checkpoints for four different models, including RIFE, IFRNet, AMT-S, and EMA-VFI.

Download checkpoints from here (full) / here (separate).

P.S., RIFE-pro denotes the RIFE model trained with more data and epochs; RIFE-vgg denotes the RIFE model trained with the LPIPS loss.

Inference

Two images

python inference_img.py --img0 [IMG_0] --img1 [IMG_1] --output_dir [OUTPUT_DIR] --model [MODEL_NAME] --variant [VARIANT] --num [NUM] --gif

Examples:

python inference_img.py --img0 ./demo/I0_0.png --img1 ./demo/I0_1.png --model RIFE --variant DR --checkpoint ./checkpoints/RIFE/DR-RIFE-pro --save_dir ./results/I0_results_DR-RIFE-pro --num 1 1 1 1 1 1 1 --gif

python inference_img.py --img0 ./demo/I0_0.png --img1 ./demo/I0_1.png --model RIFE --variant DR --checkpoint ./checkpoints/RIFE/DR-RIFE-vgg --save_dir ./results/I0_results_DR-RIFE-vgg --num 1 1 1 1 1 1 1 --gif

python inference_img.py --img0 ./demo/I0_0.png --img1 ./demo/I0_1.png --model EMA-VFI --variant DR --checkpoint ./checkpoints/EMA-VFI/DR-EMA-VFI --save_dir ./results/I0_results_DR-EMA-VFI/ --num 1 1 1 1 1 1 1 --gif

--num NUM means to interpolate NUM frames between every two frames.
--num NUM1 NUM2 ... means that NUM1 frames are interpolated between every two frames, then NUM2 frames are interpolated between every two frames for the result of the interpolation, and so on.

Video

python inference_video.py --video [VIDEO] --output_dir [OUTPUT_DIR] --model [MODEL_NAME] --variant [VARIANT] --num [NUM]

Examples:

python inference_video.py --video ./demo/demo.mp4 --model RIFE --variant DR --checkpoint ./checkpoints/RIFE/DR-RIFE-pro --save_dir ./results/demo_results_DR-RIFE-pro --num 3 --fps 15

P.S., if without --fps, the output video will have the same fps as the input video.

Manipulation

Manipulated interpolation of anything

Demos

Webapp

You can play with the interactive demo or install the webapp locally.

Install the webapp locally

P.S., not required if you use docker

Follow ./webapp/backend/README.md to setup the environment for Segment Anything.
Follow ./webapp/webapp/README.md to setup the environment for the webapp.

Run the app

cd ./webapp/backend/
python app.py

# open a new terminal
cd ./webapp/webapp/
yarn && yarn start

Dataset

You can download the splited Vimeo90K dataset with our distance indexing maps from here ( or full dataet), and then merge them:

cat vimeo_septuplet_split.zipa* > vimeo_septuplet_split.zip

Alternatively, you can download original Vimeo90K dataset from here, and then generate distance indexing (P.S. Download checkpoints for RAFT and put them under ./RAFT/models/ in advance):

python multiprocess_create_dis_index.py

Train

Training command:

python train.py --model [MODEL_NAME] --variant [VARIANT]

Examples:

python train.py --model RIFE --variant D

python train.py --model RIFE --variant DR

python train.py --model AMT-S --variant D

python train.py --model AMT-S --variant DR

Test

Testing with precomputed distance maps:

python test.py --model [MODEL_NAME] --variant [VARIANT]

Examples:

python test.py --model RIFE --variant D

python test.py --model RIFE --variant DR

Testing using uniform distance maps with the same inputs as the time indexes:

python test.py --model [MODEL_NAME] --variant [VARIANT] --uniform

Examples:

python test.py --model RIFE --variant D --uniform

python test.py --model RIFE --variant DR --uniform

Citation

If you find this repository useful, please consider citing:

@article{zhong2023clearer,
  title={Clearer Frames, Anytime: Resolving Velocity Ambiguity in Video Frame Interpolation},
  author={Zhong, Zhihang and Krishnan, Gurunandan and Sun, Xiao and Qiao, Yu and Ma, Sizhuo and Wang, Jian},
  journal={arXiv preprint arXiv:2311.08007},
  year={2023}
}

Acknowledgements

We thank Dorian Chan, Zhirong Wu, and Stephen Lin for their insightful feedback and advice. Our thanks also go to Vu An Tran for developing the web application, and to Wei Wang for coordinating the user study.

Moreover, we appreciate the following projects for releasing their code:

[CVPR 2018] The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
[ECCV 2020] RAFT: Recurrent All Pairs Field Transforms for Optical Flow
[ECCV 2022] Real-Time Intermediate Flow Estimation for Video Frame Interpolation
[CVPR 2022] IFRNet: Intermediate Feature Refine Network for Efficient Frame Interpolation
[CVPR 2023] AMT: All-Pairs Multi-Field Transforms for Efficient Frame Interpolation
[CVPR 2023] Extracting Motion and Appearance via Inter-Frame Attention for Efficient Video Frame Interpolation
[ICCV 2023] Segment Anything