LLVIP
A visible-infrared paired dataset for low-light vision tasks like pedestrian detection, image fusion, and image-to-image translation.
What is LLVIP?
LLVIP is a visible-infrared paired dataset designed for low-light vision research, containing over 30,000 aligned image pairs with pedestrian annotations. It addresses the challenge of poor visibility in dark environments by providing thermal infrared data that highlights human subjects, enabling tasks like pedestrian detection, image fusion, and cross-modal translation. The dataset serves as a benchmark for developing and evaluating algorithms that leverage multimodal imaging.
Computer vision researchers and practitioners working on low-light applications, multimodal learning, pedestrian detection, or image fusion. It's particularly relevant for those developing algorithms for surveillance, autonomous driving, or night-vision systems.
LLVIP offers a unique large-scale collection of precisely aligned visible-infrared pairs with clean annotations, filling a gap in publicly available low-light datasets. Its inclusion of baseline implementations and tools lowers the barrier to entry, while the Kaggle competition fosters community engagement and benchmarking.
Overview
LLVIP: A Visible-infrared Paired Dataset for Low-light Vision
Use Cases
Best For
- Training pedestrian detection models for low-light surveillance systems
- Researching image fusion techniques between visible and infrared modalities
- Developing image-to-image translation models for night-vision enhancement
- Benchmarking object detection algorithms in challenging lighting conditions
- Studying multimodal learning with aligned thermal and visual data
- Creating synthetic training data for low-light vision tasks
Not Ideal For
- Commercial projects requiring licensing for product deployment
- Research focused exclusively on visible-light imaging without infrared data
- Real-time applications needing lightweight, fast-processing datasets
- Studies in well-lit environments where low-light conditions are irrelevant
Pros & Cons
Pros
With 30,976 precisely aligned visible-infrared images, it offers substantial data for robust multimodal model training, as emphasized in the dataset description.
Provides implementations for image fusion, pedestrian detection, and image translation, including pre-trained models like pix2pixGAN, reducing initial setup effort for researchers.
Includes bounding box labels for pedestrians, enabling direct benchmarking for object detection in low-light scenarios, with tools for format conversion.
Hosts a Kaggle competition to foster algorithmic development, encouraging collaborative research and standardizing evaluation metrics.
Cons
The dataset is restricted to non-commercial use, which excludes industry applications and requires alternative sources for commercial projects.
Baseline implementations rely on deprecated libraries like TensorFlow 1.14.0, making environment setup complex and prone to compatibility issues.
The README notes corrections to annotation errors, indicating potential inconsistencies that researchers must account for in their work.
Frequently Asked Questions
Related Projects
Fashion-MNISTA MNIST-like fashion product database. Benchmark :point_down:
Open Images datasetThe Open Images dataset
DeepMind QA CorpusQuestion answering dataset featured in "Teaching Machines to Read and Comprehend
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.