Overview
This project generates natural language captions for satellite and aerial images using the RSICD dataset. It explores encoder-decoder approaches for converting visual scene content into meaningful textual descriptions.
Key Highlights
- Compares two modeling directions: CNN + LSTM and CNN + Transformer.
- Uses RSICD data with around 10,000 remote sensing images and 5 captions per image.
- Evaluates caption quality with BLEU and CIDEr style metrics.
- CNN + LSTM setup reaches around BLEU 0.34 and CIDEr 0.81.
Technologies
- Core: Python, PyTorch
- Modeling: Encoder-decoder caption generation
- Evaluation: BLEU, CIDEr
- Dataset: RSICD