asdmd/PVSDL_DATASET

# PVSDL: PV Soiling Detection with LLMs Dataset

### Overview
The **PVSDL (PV Soiling Detection with LLMs)** dataset is a specialized collection of images designed for detecting soiling conditions on photovoltaic (PV) panels. This dataset is specifically curated to facilitate research into leveraging Large Language Models (LLMs) and Vision-Language Models (VLMs) for intelligent industrial inspection. By accurately identifying surface contaminants, the dataset supports the development of automated systems that optimize the efficiency of solar power generation.

### Dataset Structure
The data is partitioned into three main directories to ensure a streamlined workflow for model training and unbiased performance evaluation:
- **train/**: Images used for model training.
- **val/**: Images used for hyperparameter tuning and validation.
- **test/**: Images used for final performance benchmarking.

### Labeling Convention
The dataset utilizes an efficient file-naming system where labels are integrated directly into the image titles. The naming format is `{location}_{index}_{label}.jpg`.

**Example:** `city_0001_0.jpg`
- **Label 0**: Represents a **Clean** PV panel.
- **Label 1**: Represents a **Soiled/Dirty** PV panel (e.g., dust, bird droppings, or debris).

### Core Applications
This dataset is intended for tasks including:
* **Multimodal Image Classification**: Testing the descriptive and analytical power of VLMs.
* **Zero-shot/Few-shot Learning**: Evaluating model performance on industrial inspection tasks with minimal training data.
* **Smart Maintenance**: Developing predictive cleaning schedules for renewable energy infrastructure.
---
license: mit
---