Data Filtering

The code uses OmniParser to generate bounding boxes for images, followed by a filtering step to clean noisy data.

Setup

1. Install Environment
   Follow the environment-setup instructions provided in the OmniParser repository.

2. Install Additional Dependency
   Install the imagesize Python package if it is not already included:

pip install imagesize

3. Download OmniParser Checkpoint Download the checkpoint, and place it OmniParser-v2.0 folder.

Usage

Step 1 - Generate Bounding Boxes

Run the run.py script to generate bounding boxes for all images using OmniParser:

python run.py

Step 2 - Filter Noisy Data

Run the filter.py script to clean the dataset by checking whether each data bounding box overlaps only with the OmniParser output:

python filter.py

Folder Structure

├── run.py # Script to generate bounding boxes using OmniParser
├── filter.py # Script to clean noisy datasets
├── inp.json # Raw input data
├── OmniParser-v2.0/ # OmniParser checkpoints
├── hw_cache.json # Cached height and width info for all images
├── images/ # Directory containing all images
├── log/ # Outputs from OmniParser
├── clean.json # Cleaned output data
├── README.md # This file
└── ... # Other files or directories

Notice

In inp.json, all bounding box coordinates are stored using normalized values in the range [0, 1000].