You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection

Authors: Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, Shuaicheng Liu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical findings reveal significant improvements in low-light object detection tasks, as well as promising results in both well-lit and over-lit scenarios.
Researcher Affiliation Collaboration Mingbo Hong Megvii Technology, Beijing, China mingbohong97@gmail.com Shen Cheng Megvii Technology, Beijing, China chengshen@megvii.com Haibin Huang Kuaishou Technology jackiehuanghaibin@gmail.com Haoqiang Fan Megvii Technology, Beijing, China fhq@megvii.com Shuaicheng Liu University of Electronic Science and Technology of China liushuaicheng@uestc.edu.cn
Pseudocode No The paper describes its methods verbally and mathematically but does not include any clearly labeled pseudocode blocks or algorithms in a structured format.
Open Source Code Yes Code is available at https://github.com/Mingbo Hong/YOLA.
Open Datasets Yes We evaluate our proposed method on both real-world scenarios datasets: exclusively dark [32] (Ex Dark) and UG2+DARK FACE [48].
Dataset Splits Yes Ex Dark dataset contains 7363 images ranging from lowlight environments to twilight, including 12 categories, 3,000 images for training, 1,800 images for validation, and 2,563 images for testing.
Hardware Specification Yes Specifically, we trained 12 epochs with 8 GPUs and a mini-batch of 1 per GPU in an initial learning rate of 1e-2 by the SGD optimizer on both well-lit and over-lit (generated by brightening the origin image) scenarios.
Software Dependencies No We implement YOLA using the MMDetection toolbox [4]. The paper mentions a software toolbox but does not specify its version number, nor does it list versions for other common dependencies like Python or PyTorch.
Experiment Setup Yes Both detectors are initially pre-trained on the COCO dataset and subsequently fine-tuned on the target datasets utilizing the SGD [41] optimizer with an initial learning rate of 1e-3. Specifically, we resize the Ex Dark dataset images to 608 608 and train both detectors for 24 epochs, reducing the learning rate by a factor of 10 at epochs 18 and 23.