HLGNet

High-Light Guided Network for Low-Light Instance Segmentation with Spatial-Frequency Domain Enhancement

<!-- Add framework image -->

Official implementation of HLGNet. Code will be released after publication!

Full Benchmark Results

Instance Segmentation Performance on Low-Light Datasets

| Method | APseg | AP50 | AP75 | APbox | AP50 | AP75 | APseg | AP50 | AP75 | APbox | AP50 | AP75 | |----------------|---------|--------|--------|---------|--------|--------|---------|--------|--------|---------|--------|--------| | |Belt Dataset (4,291 images) | | | | | | | Lis Dataset (2,230 images) | | | | | | | Mask RCNN | 72.9 | 82.5 | 73.0 | 78.3 | 88.8 | 80.0 | 34.2 | 55.6 | 34.7 | 41.3 | 63.9 | 44.6 | | Cascade RCNN | 73.6 | 84.4 | 73.5 | 80.0 | 90.7 | 80.3 | 34.5 | 56.2 | 35.4 | 42.5 | 66.2 | 46.1 | | DENet | 79.0 | 89.2 | 78.8 | 84.1 | 95.8 | 85.2 | 38.6 | 61.7 | 39.8 | 46.4 | 70.1 | 51.0 | | PENet | 76.4 | 85.9 | 74.2 | 81.0 | 92.8 | 82.1 | 36.1 | 58.8 | 36.4 | 43.6 | 67.3 | 47.1 | | Zero-DCE | 79.1 | 90.0 | 78.1 | 83.3 | 95.7 | 86.7 | 38.7 | 62.0 | 39.0 | 46.4 | 70.0 | 50.9 | | EnlightenGAN | 77.6 | 88.9 | 78.5 | 84.0 | 93.6 | 84.5 | 38.4 | 61.5 | 39.2 | 45.8 | 69.5 | 49.7 | | RUAS | 75.0 | 86.0 | 75.6 | 81.7 | 91.0 | 82.2 | 36.1 | 58.6 | 36.4 | 43.8 | 66.7 | 48.0 | | SCI | 76.7 | 86.9 | 75.7 | 81.8 | 92.6 | 83.9 | 36.5 | 59.5 | 37.0 | 44.3 | 67.3 | 48.4 | | NeRCo | 76.8 | 88.3 | 77.8 | 81.9 | 92.3 | 83.6 | 36.7 | 60.3 | 38.6 | 44.6 | 68.3 | 48.6 | | SMG | 77.4 | 87.2 | 77.2 | 83.3 | 92.6 | 83.4 | 37.4 | 60.3 | 38.7 | 44.7 | 67.4 | 49.2 | | YOLA | 79.3 | 89.8 | 80.0 | 85.4 | 95.8 | 87.1 | 39.8 | 63.5 | 41.4 | 47.5 | 70.9 | 51.8 | | Mask2former | 74.4 | 83.0 | 73.3 | 76.3 | 81.1 | 75.3 | 35.6 | 55.2 | 35.2 | 37.8 | 55.9 | 39.9 | | YOLOv8-seg | 73.1 | 82.6 | 73.4 | 82.4 | 88.6 | 83.9 | 34.3 | 56.0 | 34.9 | 45.1 | 64.3 | 48.3 | | PointRend | 71.5 | 80.8 | 78.1 | 75.7 | 82.9 | 75.0 | 32.8 | 52.9 | 39.8 | 37.1 | 57.9 | 39.8 | | Lis | 79.8 | 90.6 | 80.5 | 86.3 | 93.4 | 86.5 | 40.8 | 62.7 | 41.5 | 48.0 | 69.2 | 52.6 | | HLGNet (Ours) | 81.6 | 92.8 | 82.3 | 88.1 | 97.4 | 88.5 | 42.2 | 65.5 | 43.6 | 50.3 | 72.4 | 53.7 |

All values are percentages. Higher is better. Highlighted values indicate state-of-the-art performance.

Dataset Preparation

1. Belt Dataset

Dataset included in repo:
/belt/ directory contains full dataset (4,291 images with annotations)

2. Lis Dataset

Download from official source:
https://github.com/Linwei-Chen/LIS

Usage (Coming Soon)

```bash

Installation

git clone https://github.com/yourusername/HLGNet.git cd HLGNet pip install -r requirements.txt

Training

python train.py --dataset belt --config configs/base.yaml

Evaluation

python test.py --checkpoint pretrained/hlgnet.pth --dataset lis