Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network
Authors: Fangtong Sun, Congyu Li, Ke Yang, Yuchen Pan, Hanwen Yu, Xichuan Zhang, Yiying Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across various downstream tasks demonstrate that FRBNet achieves superior performance, including +2.2 m AP for dark object detection and +2.9 m Io U for nighttime segmentation. We conduct extensive experiments to evaluate the effectiveness of the proposed plug-and-play FRBNet on low-light vision tasks of detection and segmentation. Specifically, we adopt Ex Dark[40], Dark Face[73], ACDC-night[50], and LIS[4] datasets for dark object detection, face detection, nighttime semantic segmentation, and dark instance segmentation tasks, respectively. |
| Researcher Affiliation | Academia | EMAIL EMAIL, EMAIL |
| Pseudocode | No | The paper describes the FRBNet architecture and its components (Illumination-invariant feature enhancement process in the frequency domain, Learnable frequency-domain filter) using mathematical formulations and descriptive text, but it does not include any clearly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Code is available at: https://github.com/Sing-Forevet/FRBNet. |
| Open Datasets | Yes | Specifically, we adopt Ex Dark[40], Dark Face[73], ACDC-night[50], and LIS[4] datasets for dark object detection, face detection, nighttime semantic segmentation, and dark instance segmentation tasks, respectively. |
| Dataset Splits | Yes | Consistent with the official experimental setup in UG2+ Challenge, we adopt a 3:1:1 random split of the Dark Face dataset for training, validation, and testing in our experiments. Table 6 summarizes the statistics of our employed datasets. These datasets cover a wide range of low-light vision tasks, including object detection, face detection, semantic segmentation, and instance segmentation. # Class is the number of classes, whereas #Train and #Val denote the number of training and validation samples for each dataset, respectively. |
| Hardware Specification | Yes | Experiments are implemented based on the MMDetection [2] and MMSegmentation [6] toolboxes by Py Torch and trained on a NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | Experiments are implemented based on the MMDetection [2] and MMSegmentation [6] toolboxes by Py Torch and trained on a NVIDIA RTX 4090 GPU. |
| Experiment Setup | Yes | Following the experimental setup of YOLA [5], we set the momentum and weight decay of the SGD optimizer for the detection model to 0.9 and 0.0005, respectively. The learning rate is 0.001. For Ex Dark, all input images are resized to 608 608, and both detectors are trained for 24 epochs. For Dark Face, YOLOv3 maintains 608 608 and is trained for 20 epochs, while TOOD uses a higher resolution of 1500 1000 and is trained for 12 epochs. |