frameintIFE

KO | EN

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #deepseek-harness #dsh-plugin #open-source #ai #cli #claude-code #windows #codex #developer-tools #dsh

★ 41 Open GitHub ↗

Official PyTorch implementation of "Frame Interpolation for Dynamic Scenes with Implicit Flow Encoding"

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Frame Interpolation for Dynamic Scenes with Implicit Flow Encoding

### [Project](https://people.engr.tamu.edu/nimak/Papers/WACV2023_Interp) | [Paper](https://arxiv.org/pdf/2209.13284) | [Supplementary](https://people.engr.tamu.edu/nimak/Papers/WACV2023_Interp/media/WACV2023_Interp_supp.pdf) | [YouTube](https://youtu.be/Re_c-CBlSfI)

Official PyTorch implementation of our frame interpolation architecture. We leverage pre-trained [RAFT](https://github.com/princeton-vl/RAFT) to generate the flows we use as input to our Flow Decoder. Then, we combine the two warped images and their features using [FILM](https://github.com/google-research/frame-interpolation)'s pre-trained blending network. Our method is robust to lighting variation, being a per-scene optimization alternative to challenging dynamic scenes.

[Frame Interpolation for Dynamic Scenes with Implicit Flow Encoding](https://arxiv.org/abs/2209.13284) 
[Pedro Figueirêdo](http://people.tamu.edu/~pedrofigueiredo/), [Avinash Paliwal](http://people.tamu.edu/~avinashpaliwal/), [Nima Khademi Kalantari](https://people.engr.tamu.edu/nimak/index.html) 
Texas A&M University 
In WACV 2023

![A sample 2 seconds moment.](https://raw.githubusercontent.com/pedrovfigueiredo/frameintIFE/main/misc/baby.gif)
FrameintIFE interpolates challenging near-duplicate photos, creating a slow motion video that depicts the natural transition between them.

## Installation

### We have tested this project with Ubuntu 22.04 and Python 3.9.13.

*   Clone this repository and its submodules

```
git clone --recurse-submodules https://github.com/pedrovfigueiredo/frameintIFE frameintIFE
cd frameintIFE
```

*   Set up an [anaconda](https://www.anaconda.com/) environment
```
conda create -n frameintIFE python=3.9.13
conda activate frameintIFE
```

*   Install Tensorflow and its dependencies
```
conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/
pip install -r requirements-tf.txt
```

*   Install Pytorch and its dependencies
```
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
conda install opencv=4.6.0 tensorboardx=2.2 -c conda-forge
```

*   Apply changes to submodule pytorchmeta before installing it

```
cd lib/pytorch-meta
git checkout d55d89ebd47f340180267106bde3e4b723f23762
git apply ../../diffs/pytorch-meta.diff
python setup.py install
```

## Pre-trained Models

* We use the [sintel checkpoint for RAFT](https://drive.google.com/drive/folders/1sWDsfuZ3Up38EUQt7-JDTT1HcGHuJgvT?usp=sharing) and the [style checkpoint for FILM](https://drive.google.com/drive/folders/1q8110-qp225asX3DQvZnfLfJPkCHmDpy?usp=sharing). We also provide [pre-trained weights of the Flow Decoder](https://drive.google.com/drive/folders/1cHg5Qj0g08I6jkQTkhjXii4qbeZmwCLC?usp=share_link) for the sample data.
* Download the required weights from each google drive source, placing them into a newly created directory `<pretrained_models>`.

## Running the Codes

The following instructions demonstrate the frame interpolation of sample data stored in frameintIFE/sample_data.

### Generating RAFT reference flows

RAFT requires significant GPU memory for higher resolution images. Choose the `<factor>` argument to adjust the sample data resolution (originally 4K) according to your memory allowance. A factor of 0.25 is viable with a Nvidia RTX 3080 (10GB). For the paper, we use a factor of 0.5 on a Nvidia A100.

```
python RAFT/generateGTs.py --model <pretrained_models>/raft-sintel.pth --path sample_data/baby --output_path sample_data_out/baby --factor 0.25
```
Flows and images will be placed into `<output_path>`.

### Optimizing Flow Decoder

We provide [pre-trained weights for the Flow Decoder](https://drive.google.com/drive/folders/1cHg5Qj0g08I6jkQTkhjXii4qbeZmwCLC?usp=share_link) for the sample data scenes. Alternatively, you may optimize the flow decoder yourself by following the instructions below.

The number of optimization steps needed for a high-quality interpolation vary depending on the scene and its resolution. For all scenes shown on the paper and supplementary video, we use 10K iterations at a learning rate of 1e-6.

Optionally, you may set a `<savepath>` to get intermediate evaluations during the optimization, including flows and warped frames.

Execute the following command to optimize the Flow decoder, getting a continous representation of the RAFT flows generated on the previous step.

```
python optimize_decoder.py --f_left <output_path>/f42.pt --f_right <output_path>/f24.pt --siren_usebias --hyp_usebias --save_on_flow_dir
```

### Generating final frames

To generate final blended frames, we first query our continous flow representation at intermediate $t \in [0.0, 1.0]$. To match FILM's recursive interpolation timestamps, we use a `<interp_size>` argument, as follows:

```
python discretize.py --dir <output_path> --interp_size 5 --siren_usebias --hyp_usebias
```

However, you can modify the input $t$ in the discretization code to generate any timestamp. This step generates a `flows.npy` file containing discretized flow values under `<output_path>`.

Now, we can use FILM's blending network to generate the final frames and accompanying video.

Run the following command to execute a modified version of FILM's blending network which uses our flows. Optionally, you may generate FILM's original results by omitting the `insert_flows` option. You may also add the `output_detailed` option to output both warped images and flows.

```
python -m FILM.eval.interpolator_cli --pattern <output_path> --model_path <pretrained_models>/film_net/Style/saved_model --times_to_interpolate 5 --insert_flows --output_video
```

Note: FILM's results worsen as the image resolution increases. Therefore, when comparing lower resolution interpolations (e.g. 1K resolution as a result of a factor of 0.25 during RAFT pre-processing), you may see less obvious artifacts than the reported in the paper. We suggest a per-frame comparison in these cases.

## Supplementary Video
[![Supplementary video link to youtube](https://raw.githubusercontent.com/pedrovfigueiredo/frameintIFE/main/misc/thumbnail.jpg)](https://youtu.be/Re_c-CBlSfI)

## Citation
```
@InProceedings{Figueiredo_2023_WACV,
    author    = {Figueir\^edo, Pedro and Paliwal, Avinash and Kalantari, Nima Khademi},
    title     = {Frame Interpolation for Dynamic Scenes With Implicit Flow Encoding},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {218-228}
}
```

## Acknowledgments

We would like to thank Ben Figueiredo, Carla Figueiredo, and Fernanda Gama for their contribution with the qualitative scenes.

frameintIFE

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README