# evaluate

**Repository Path**: Im34v/evaluate

## Basic Information

- **Project Name**: evaluate
- **Description**: A study of metrics for assessment on videos of invisible light.
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-07-12
- **Last Updated**: 2023-02-13

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# evaluate

  This evaluate repository is primarily a reference to three GitHub projects, [deepmind](https://github.com/deepmind)/**[kinetics-i3d](https://github.com/deepmind/kinetics-i3d)**, [hassony2](https://github.com/hassony2)/**[kinetics_i3d_pytorch](https://github.com/hassony2/kinetics_i3d_pytorch)** and [miracleyoo](https://github.com/miracleyoo)/**[Trainable-i3d-pytorch](https://github.com/miracleyoo/Trainable-i3d-pytorch)**.

## dataset

### data directory

```
├── images // If you use image series instead of video as raw input
│   ├── pre-processed
│   │   ├── class1
│   │   │   └── Series1
│   │   │   │   ├── rgb-SampleNum_xx.npy
│   │   │   │   └── flow-SampleNum_xx.npy
│   │   │   ├── ...
│   │   └── ...
│   └── raw
│       ├── class1
│       │   └── Series1
│       │       ├── pic1.png
│       │       ├── pic2.jpg
│       │       ├── pic3.jpg
│       │       ├── ...
│       │       └── picx.jpg
│       │   ├── ...
│       └── ...
└── videos // If you use video as raw input
    ├── pre-processed
    │   ├── train
    │   │   ├── class1
    │   │   │   └── Series1
    │   │   │       ├── rgb-FPS_xx.npy
    │   │   │       └── flow-FPS_xx.npy
	│   │   │   ├── ...
	│   │   └── ...
    │   └── val
    │       ├── class1
    │       │   └── Series1
    │       │       ├── rgb-FPS_xx.npy
    │       │       └── flow-FPS_xx.npy
	│       │   ├── ...
	│       └── ...
    └── raw
        ├── class1
        │   ├── Series1.mp4
        │   ├── Series2.mp4
        │   └── ...
        └── ...
```

### Kinetics-400

  You may need the Kinetics-400 during testing, if you don't happen to have this dataset you can download it by clicking [here](https://pan.baidu.com/s/1ISqYffUJO53jaFdQuzZ0aA)[password: wwwe]. After downloading you will see multiple `compress.tar.gz.xx` files in raw-part, then

```
# merging segmented documents
cat compress.tar.gz.* >> compress.tar.gz
# unzip the file
tar xvf compress.tar.gz
```

## how to run

### pre-process

  You can use pictures or videos as pre-processing input. 

```
mass: Compute RGBs and Flows massively.
init_dir: Initialize the data pre-processed folder tree.
input_path: Path to input video or images folder
out_path: Where you want to save the output rgb and flow files.
resize: Resize the size of pictures.
random_choice: Whether to choose frames randomly or uniformly.
rgb: Whether to generate rbg data.
flow: Whether to generate flow data.
```

  If you use pictures as input, you had better pay attention to the parameters:

```
is_image: Use a series of images(its folder) as input.
sample_num: The number of the output frames after the sample.
sample_type: Choose the sample method. If you choose images, the value should be set to "num".
For example:
python pre_process.py --rgb --is_image --sample_num=xx --sample_type=num --resize=xx(default: 224) --input_path=xxx
```

  If you use videos as input directly, you had better pay attention to the parameters:

```
in_fps: Video's FPS can be directly acquired. The default value is 30.
out_fps: The FPS of the output video. The default value is 5.
At this time: sample_num = frames_sum * out_fps / in_fps
For example:
python pre_process.py --rgb --sample_type=fps --input_path=xx
```

### predict

  This part is relatively simple, look at the `predict.py` code to understand how to predict.

### retrain

  To be continue ...

### compute FVD

  For convenience, you can put the video pairs for which you want to calculate the FVD value under the directory `evaluate\i3d\data\videos\raw\test`. In the directory `test`, name the folder with the original video name and put the original video and the generated video under that folder. The specific file directory is shown below.

```
├── origin1
│   ├── origin1.mp4
│   ├── blur_1.mp4
│   └── ...
├── origin2
│   ├── origin2.mp4
│   ├── blur_2.mp4
│   └── ...
└── ...
```