DGFF: Detection-Guided Feature Feedback
Model repository for "End-to-End Task-Oriented Low-Light Image Enhancement via Detection-Guided Feature Feedback" (ICCA 2026).
The detector teaches the enhancer what to preserve, and the gated adapters learn to pass only the useful lessons.
| Resource | Link |
|---|---|
| Code | https://github.com/ShuvroSankar/DGFF |
| Data (LOL + ExDark) | https://huggingface.co/datasets/abbaab/DGFF-dataset |
About DGFF
DGFF is a training framework that routes intermediate YOLOv8n backbone features (P3, P4, P5) back into the decoder of a U-Net low-light enhancer (LLEN) through lightweight gated 1x1 adapters. The detector's view of what matters shapes the reconstruction, so enhancement preserves edges and textures that detectors rely on. The adapters are discarded at inference, so DGFF adds zero overhead over a plain LLEN + detector pipeline.
| Component | Parameters |
|---|---|
| LLEN (U-Net) | 4,844,803 |
| DGFF adapters (training only) | 86,464 |
| YOLOv8n backbone (frozen) | 1,272,656 |
Headline results
| Benchmark | Metric | Result |
|---|---|---|
| LOL (validation) | PSNR / SSIM | 19.83 dB / 0.9048 |
| ExDark (domain-matched) | mAP@0.5 | 0.474 (raw baseline 0.408, LLEN-only 0.469) |
| ExDark (domain-matched) | mAP@0.5:0.95 | 0.280 (raw baseline 0.232, LLEN-only 0.276) |
The +0.5-point mAP@0.5 gain over LLEN-only is directionally consistent but not statistically significant (paired bootstrap, 95% CI [-0.0065, 0.0112]). The +6.6-point gain over the unenhanced baseline is clear. Full details are in the paper and the GitHub README.
Contents
| File | Description |
|---|---|
README.md |
This card |
| (to add) | LLEN checkpoint after Phase 1 (LOL, best epoch 185) |
| (to add) | DGFF adapter weights after Phase 2 (ExDark fine-tuning) |
How to use
- Download the weights from this repository (
huggingface-cli download abbaab/DGFF --local-dir ./weights). - Clone the code:
git clone https://github.com/ShuvroSankar/DGFF.git - Download the data from abbaab/DGFF-dataset.
- Follow the training and evaluation steps in the GitHub README.
At inference only LLEN is needed: enhance the image, then run any detector (the paper uses YOLOv8n).
Limitations
- Trained on LOL (indoor, 485 pairs) and fine-tuned on ExDark, so an indoor/outdoor domain gap remains.
- Evaluated with YOLOv8n only; other detectors are untested.
- Classes with small, occluded or homogeneous-texture objects (Chair, Motorbike, People, Table) did not improve over LLEN-only.
Citation
@inproceedings{sen2026dgff,
title = {End-to-End Task-Oriented Low-Light Image Enhancement via Detection-Guided Feature Feedback},
author = {Sen, Shuvro Sankar and Mia, MD. Maruf and Hasan, Naim and Chayon, Muhammad Hasibur Rashid},
booktitle = {Proceedings of The 4th International Conference on Computing Advancements (ICCA 2026)},
year = {2026},
address = {Dhaka, Bangladesh},
publisher = {ACM}
}
Authors
Shuvro Sankar Sen, MD. Maruf Mia, Naim Hasan, Muhammad Hasibur Rashid Chayon, American International University - Bangladesh (AIUB).
Acknowledgements
Trained and evaluated on LOL (Wei et al., 2018) and ExDark (Loh and Chan, 2019), using Ultralytics YOLOv8. Dataset terms of the original sources apply.