InnoColor

InnoColor deployment for Ubuntu. PyTorch and TKinter. If you want the stable production version, please move to the release/windows branch.

innocolor/
│── main.py
│── settings.py: app and device (Linux and XBox) specifications
|── base/: parent objects for utils
|── utils/: interface objects and also app-specific processes for running models
|── interfaces/: script for the three TK interfaces
|── icons/
|
|── dsX.pth: datasets saved with information for each input-groundtruth pair along with model used
|
|── model/
|── checkpoint/
|── cvdsim/
|── sam2/
|── sam2_configs/
|
│── requirements.txt
│── README.md

Installation

1. Install Repo

git clone https://github.com/Brocoliman/InnoColor.git
cd InnoColor

2. (Recommended) Create Conda environment This project was developed and tested using Conda on Linux (Ubuntu).

conda create -n innocolor python=3.11
conda activate innocolor

3. Install requirements

pip install -r requirements.txt

3. Download models (move these into their respective folders in checkpoint/) from Google Drive

• Dalt-NET requires classifier_190.pth and LUTs_190.pth
• InnoColor requires classifier_2000.pth, generator_200.pth, LUTs_2000.pth, and SAM2UNet-SOD.pth
• distill requires generator_100.pth

4. Download sample datasets and move them into root folder

• ds1.pth
• ds2.pth
• ds3.pth

5. Change FOLDER in settings.py to the name of this project location

Run

python main.py

Quick Guide

Each interface (camera interface and data-editing interface) contains components of the InnoColor toolbar. This includes the model, sim, and highlight selection.

You can choose between model types of

• None: No Daltonization is applied
• Hue Rotation [P]: JB Huang's Hue Rotation matrix as described in Huang et al., 2008
• Dalt-NET [P]: multi-mapping table based method described in Wang, 2024
• InnoColor [P]: saliency based method described in Wang, 2025. Note that this is NOT the newest method
• Distillation [D]: newest method using data distillation into DDPA. Please read the attached paper
• Usually models developed for protanopia are fairly suitable for deuteranopia cases, and vice versa, although there are significant limitations.
• In the data-editing interface, models have an input parameter slider. No current model utilizes this yet, although this value is passed as a parameter

You can choose between sim types of

• None: no CVD
• Protan: protanopia or protanomaly, red cone deficiency, second most common case
• Deutan: deuteranopia or deuteranomaly, green cone deficiency, most common case
• Tritan: tritanopia or tritanomaly, blue cone deficiency, a much rarer case
• Both interfaces allow slider adjustment of the severity from 0 (no CVD) to 1 (-nopia, full blindness for respective cone)

You can highlight pixels of colors near

• None
• Red
• Green
• Blue