LLEN-SLAM

基于弱光增强的SLAM

💡 News 新闻

• 2024.10.03 V2.3 on Epoch 870/1500, have PSNR: 26087951, SSIM: 0.793044 on SID
• 2024.10.03 For the SID, SMID, SDSD data sets, we reworked these data sets.
• 2024.09.26 V2.3 on Epoch 370/1500, have PSNR: 28.329210, SSIM: 0.877992 on LOL_v2
• 2024.09.23 V2.1.1 version in epoch150, PSNR:25.5085, SSIM:0.8311 on LOL_v1
• 2024.09.21 V2.0-6variate version in epoch150, PSNR:25.429, SSIM:0.8292
• 2024.09.18 V1.5 version in epoch150, PSNR:25.39
• 2024.09.15 V1.4 version in epoch150, PSNR:24.97, SSIM:0.811
• 2024.09.03 🌟 The first version completed training and testing, PSNR: 25.441431, SSIM: 0.824064. The input conversion to LAB is done in this version. Good results were achieved.
• 2024.08.19 The initial model was completed, and compared with different low-light models, the effect was only 20.9. Other models were tested simultaneously (avg:21.5), with a difference of 0.6 points.
• 2024.08.08 project startup 🎈

⚙ module 模型组件

🖼 Visual Comparison 视觉比较

🧾 Weights and Results

Ours

Folder (test datasets)	PSNR	SSIM	LPIPS	Results	Weights Path
(LOLv1) v1	25.441431	0.824064	**	**
(LOLv2) v2	28.329210	0.877992	**	**

Compare LOLv1 datasets

It should be noted that the following assessment is not fine-tuned and is conducted directly after training.

Model name	PSNR	SSIM	LPIPS	Weights Path
Ours_v1	25.441431	0.824064	X	LOLv1/net_g_1500.pth
HVI-CIDNet	21.61	0.793	0.217	LOLv1\My_model\epoch_best.pth
retinexformer	21.798	0.802	X	LOLv1\My_model\best_psnr_21.96_27000.pth
URetinex-Net	21.32(Official)	X	X	LOLv1\Official_model\ckpt
GASD	26.232	0.8519	X	LOLv1\My_model
LYT—NET	25.4462	0.8307	X	LOLv1\My_model
KinD++	15.6506	0.5079	0.2555	LOLv1\My_model
Diffusion Low Light	21.651	0.8056	0.1768	LOLv1\My_model

You can get pth on Google drive

Compare LOLv1 datasets official numerical value

Model name	PSNR	SSIM	LPIPS
Ours_v1	20.9049	0.7718	X
HVI-CIDNet	23.500	0.8703	0.1053
retinexformer	25.154	0.8445	X
URetinex-Net	21.32	X	X
GSAD	27.623	0.874	0.0912
LYT—NET	26.6280	0.8349	X
KinD++	X	X	X
Diffusion Low Light	26.047	0.8445	0.1184

Training log

Model name	comment
Ours_v1	epoch:1500
Ours_v2	epoch:1500
HVI-CIDNet	epoch:1000
retinexformer	epoch:683 & iter:41K
URetinex-Net	-
GSAD	First training 1000K;Second training 2000K
LYT—NET	epoch:1500
KinD++	-

check log

Log files are stored in ./Compare_models/ different models/ LOLv1/info

🌑 0. My environment

• Ubuntu 20.04.6
• AMD R9 5900HX
• RTX 3080 Laptop 16G
• RAM 32G

🌑 1. Get Started

• Python 3.9
• pytorch-cuda=11.8

(1) Create Conda Environment

conda create --name LLEN python=3.9 -y
conda activate LLEN

(2) Clone Repo

git clone https://github.com/suiuko/LLEN-SLAM.git

(3) Install Dependencies

cd LLEN-SLAM
pip install -r requirements.txt

Data Preparation

Currently only the LOLv1 database is tested

• LOLv1

Note: (1) For the SID, SMID, SDSD data sets, we reworked these data sets. Convert the .npy file to an RGB file, which can be downloaded if you want to use the reconstructed data set.

datasets (click to expand)

├── datasets
	├── DICM
	├── LIME
	├── LOLdataset
		├── Train
			├──low
			├──high
		├── Test
			├──low
			├──high

🌒 2. Testing

🌒 3. Training

# activate the enviroment
conda activate LLEN

# LOL-v1
python3 train.py