CS7180 LLE Adaptive Curve Estimation

Reimplements this paper https://arxiv.org/abs/2308.08197

Teammates: Richard Zhao(Training), Oliver Fritsche(Testing), Yunpei Gu (model) Course: CS 7180 Advanced Perception
Operating System: Linux 64, MacOS

Project Overview

This project implements Stage1 of the Self-Reference Deep Adaptive Curve Estimation (Self-DACE) method for low-light image enhancement. The implementation is based on the Zero-DCE architecture and focuses on learning pixel-wise curve parameters for image enhancement.

Data

Following the paper, we train using

For training, we used SICE Part 1 dataset (2002 images), which is provided in https://github.com/Developer-Zer0/ZeroDCE/tree/main/Dataset

For testing and evaluation, we used:

• Model Evaluation with Lol eval15 set: https://www.kaggle.com/datasets/soumikrakshit/lol-dataset
• Qualitative comparison used images from https://github.com/mahmoudnafifi/C5/tree/main/images and from author's repository https://github.com/John-Wendell/Self-DACE/blob/main/visualization/data

Environment / Dependencies

To install dependencies

pip install -r requirements.txt

We used Python >3.10 and dependencies include:

• pytorch
• torchvision
• numpy
• matplotlib
• pillow
• jupyter / ipykernel
• tqdm

How to run

Option A — Notebook (for specific figures)

• Store your low-light images (recommend to use png format) in data/random/low_num

• (Optional) Store your high-light images in data/random/high_num

• Launch Jupyter and run visualization_and_evaluation_special.ipynb.

• Run the Config & model load cell (it expects epoch_118_model.pth at repo root).

• Run the Batch evaluation cell:

  Enhances all images in data/random/low_num.

  If a same-number GT exists in high_num, computes PSNR/SSIM.

  Saves triptychs to outputs/random/triptychs_num/.

• (Optional) Run the Iteration visualization cell to export “Original → Iter1..7 → Final”.

• The final Composite cell exports a single figure with Original | Self-DACE | GT rows. -->

Option B — Notebook (for LOLdataset)

Open and run visualization_and_evaluation.ipynb

Model

Model is in model.py

quick sanity test of model.py:

python model_test.py

Expected: prints shapes like:

enhanced: torch.Size([1, 3, 256, 256]) ...
alpha_stack: torch.Size([1, 21, 256, 256])
beta_stack : torch.Size([1, 21, 256, 256])

Used in training:

import torch
import model  # model.py

net = model.light_net().cuda().train()
# x0 in [0,1], shape Bx3xHxW
enhanced, alpha_stack, beta_stack = net(x0)

Training

python train.py

Run tensorboard only if you're training locally tensorboard --logdir=runs

All hyperparameters are hardcoded as constants in that file.