# Clothes Segmentation

![Sample Images and Segmentation Masks from Dataset](assets/dataset_examples.png)

This project provides a solution for segmenting clothes into 18 categories using DINO, ViT and UNet models.

* DINO: Pretrained backbone with a segmentation head
  * https://arxiv.org/abs/2104.14294
  * https://huggingface.co/facebook/dinov2-small
* ViT: Pretrained vision transformer with a segmentation head
  * https://arxiv.org/abs/2010.11929
  * https://huggingface.co/google/vit-base-patch16-224
* UNet: Custom implementation
  * https://arxiv.org/abs/1505.04597

Gradio is used for building a web interface and Weights & Biases for experiments tracking.

## Installation

1. Clone the repository:
    ```bash
    git clone https://github.com/your-project/clothes-segmentation.git
    cd plant-classifier
    ```

2. Create and activate a virtual environment:
    ```bash
    python -m venv venv
    source venv/bin/activate
    ```

3. Install dependencies:
    ```bash
    pip install -r requirements.txt
    ```

## Usage

### Training the Model
   To train a model, specify one of the following using the --model argument: **dino**, **vit** or **unet**.
   ```bash
   python src/train.py --model dino
   python src/train.py --model vit
   python src/train.py --model unet
   ```

   You can also adjust other parameters, such as the number of epochs, batch size, and learning rate, by adding additional arguments. For example:
   ```bash
   python src/train.py --model unet --num-epochs 20 --batch-size 16 --learning-rate 0.001
   ```

### Launching the Gradio Interface
   ```bash
   python app.py
   ```

   Once the interface is running, you can select a model, upload an image and view the segmentation mask.
   
   ![Web Interface Screen](assets/spaces_screen.jpg)

   #### добавить ссылку

## Results

| Model      | Test Micro Recall | Test Micro Precision | Test Macro Precision | Test Macro Recall | Test Accuracy | Test Loss | Train Micro Recall | Train Micro Precision | Train Macro Precision | Train Macro Recall | Train Accuracy | Train Loss |
|------------|-------------------|----------------------|----------------------|-------------------|---------------|-----------|--------------------|-----------------------|-----------------------|--------------------|----------------|------------|
| DINO       | 0.94986          | 0.94986             | 0.71364             | 0.67052          | 0.94986       | 0.53124   | 0.97019            | 0.97019              | 0.78185              | 0.72336           | 0.97019        | 0.30441    |
| ViT        | 0.9358           | 0.9358              | 0.63939             | 0.58365          | 0.9358        | 0.71193   | 0.96734            | 0.96734              | 0.74418              | 0.66295           | 0.96734        | 0.31166    |
| UNet       | 0.95798          | 0.95798             | 0.76354             | 0.7289           | 0.95798       | 0.56764   | 0.98035            | 0.98035              | 0.82934              | 0.82688           | 0.98035        | 0.25301    |

### Training Results of DINO

![DINO_test](assets/dino_test_plots.png)

![DINO_train](assets/dino_train_plots.png)

### Training Results of ViT

![ViT_test](assets/vit_test_plots.png)

![ViT_train](assets/vit_train_plots.png)

### Training Results of UNet

![UNet_test](assets/unet_test_plots.png)

![UNet_train](assets/unet_train_plots.png)