# Artifact Images Without Background

This dataset contains images of **archaeological artifacts with the background removed**.  
The images were generated by filtering and processing the original dataset:

https://huggingface.co/datasets/archaia/dataset_v1

The following sections describe the processing pipeline used to obtain the final dataset, including:

1. Image filtering using **CLIP**
2. **Object detection and background removal**

---

# Processing Pipeline

## 1. Image Filtering with CLIP

The first step filters the images from the original dataset to retain only those containing archaeological artifacts.

The filtering process is implemented in [`ComputeCls.py`](ComputeCls.py).

### Execution

Run the following command:

```bash 
python ComputeCls.py --input_path [path_to_input_dir]
```

### Output

The script performs the following operations:

- Classifies images using CLIP
- Generates a JSON file containing classification results
- Moves images classified as artifacts to a new directory
- Saves visualization plots of images identified as artifacts
- This step produces a filtered subset containing only artifact-related images.

## 2. Object Detection and Background Removal

The second stage detects artifacts in the images and removes their backgrounds.

The code is implemented in [`test_finetuned_removebg.py`](test_finetuned_removebg.py).

### Execution

Run the following command:

```bash 
python test_finetuned_removebg.py --input_path [path_to_input_dir] --output_path [path_to_output_dir]
```

### Processing Steps

- The script iterates over each image in the input directory.
- An object detection model detects potential artifacts in the image.
- The detection model is a fine-tuned DETR model trained on 303 labeled images.
Details about the training data and the model checkpoint will be provided in a separate repository.
- Only detections labeled as "*Archaeological artifacts*" are retained.
- The detected bounding boxes are used to crop the original images.
- Since a single image may contain multiple artifacts, each detection is processed independently.
- For every cropped artifact, the BRIAAI background removal model is applied to remove the background.

### Output

- The processed images are saved in the specified output directory.
- If multiple artifacts are detected in the same image, each artifact is saved separately using the following naming convention: original_filename_[detection_id].png

    - Example:
        - 000000029_0.png
        - 000000029_1.png
        - 000000029_2.png

- Each output image contains a single archaeological artifact with the background removed.

# Result

The final dataset consists of:

- Images containing individual archaeological artifacts
- Background removed
- Artifacts automatically extracted from the original dataset using a pipeline based on:
    - CLIP image filtering
    - DETR object detection
    - BRIAAI background removal