MIRNet finetuned on the custom dataset for 100 epochs

Model weights for the MIRNet implemented in this repository.
The model has been trained for 100 epochs on the LoL dataset scaled down to 64x64 images (based on the weights released here), and then fine-tuned for 100 epochs on the custom dataset released here, with early stopping intervening at 35 epochs.
It and obtains a PSNR of 19.64 on the custom-dataset test data, and increase of 1.24 in comparison to the PSNR of 18.4 obtained by the pre-trained model. With the fine-tuned model, the enhanced images contain less noises, and have colors that better correspond to the ground truth.

The performance on the custom test data can be visualized in the following figure:

Fine-tuning event photography dataset

This archive contains the fine-tuning dataset entirely composed of photographs taken by myself. The archive contains three folders, train, val and test. Each has an imgs and a targets folder containing the darkened and real images respectively.

To use this data with this codebase, simply extract the archive and move the three folders to the data/finetuning.
Reminder: as for the finetuning data, it can be downloaded and processed by running dataset_generation/pretraining_generation.py.

MIRNet pre-trained on LoL for 100 epochs

Model weights for the MIRNet implemented in this repository.
The model has been trained for 100 epochs on the LoL dataset scaled down to 64x64 images. It and obtains a PSNR of 19.5 on the LoL test data, and of 18.57 on the testing data of the custom fine-tuning dataset generated in this project (without fine-tuning).

The performance on the LoL test data can be visualized in the following figure: