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..

End-to-End Low-Light Enhancement for Object Detection with Learned Metadata from RAWs

Authors: Xuelin Shen, Haifeng Jiao, Yitong Wang, Yulin HE, Wenhan Yang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We implement our CRM-IR scheme on various object detection networks, and extensive experiments under low-light conditions demonstrate that it can significantly improve performance with an additional bitrate cost of less than 10 3 bits per pixel.
Researcher Affiliation Academia 1Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) 2Peng Cheng Laboratory 3College of Computer Science and Software Engineering, Shenzhen University EMAIL, EMAIL, EMAIL EMAIL, EMAIL
Pseudocode No The paper describes methods using mathematical equations and figures for overall framework and module details, but does not present structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/haifengjiao001/CRM-IR.
Open Datasets Yes Additionally, we present the Raw in Dark (RID) dataset, containing 500 annotated RAW sensor pairs from low-light daily scenes, further advancing RAW-based object detection. [...] Our work introduces new methods and datasets, while have alredy been released.
Dataset Splits Yes Low-light Object-Detection (LOD) dataset [42] is employed as our benchmark, which contains 2230 paired RAW and s RGB-JPEG format image pairs collected by a Canon EOS 5D Mark IV camera, covering both low-light and normal daylight scenes, where only the low-light parts are selected in our experiments. In particular, 1,784 training pairs and 446 test pairs are selected for model training and evaluation, respectively.
Hardware Specification Yes All experiments were conducted using Py Torch on an NVIDIA RTX 6000 Ada Generation GPU with 48 GB of memory.
Software Dependencies No The paper mentions 'Py Torch' as the framework used, but does not provide specific version numbers for PyTorch or any other software dependencies such as Python, CUDA, or specific libraries.
Experiment Setup Yes During training, data augmentation strategies are employed, including random horizontal flips and random scale jitter during resizing. All models were trained for 300 epochs using the Adam optimizer [47]. A linear scaling learning rate with a cosine decay schedule was employed, starting from an initial learning rate of 5e-4. The weight decay was set to 0, momentum was 0.9 and the training batch size was set to 8. During both training and testing, all input images were resized to 512 512.