luisml77/aeriald_o3_500

Gemma3 O3-500 Dataset

500 high-quality OpenAI o3-enhanced referring expressions used for distilling Gemma3-Aerial-12B, the LLM used for generating over 1 million referring expressions in the Aerial-D dataset.

Links

• 📝 Paper - arXiv preprint
• 🔮 Distilled Model - Gemma3-Aerial-12B
• 📊 Aerial-D Dataset - Aerial-D (1.5M+ expressions)
• 💻 Code - Generation pipeline
• 🌐 Project - Project page

Overview

• Size: 500 samples
• Source: Rule-based expressions from Aerial-D
• Enhancement: OpenAI o3 with dual-image prompting
• Purpose: Training data for Gemma3 distillation

Structure

enhanced_annotations_o3_dual/
├── {PATCH_ID}_group_{GROUP_ID}/
│   ├── {PATCH_ID}_clean.png        # Original aerial image
│   ├── {PATCH_ID}_mask.png         # Mask overlay
│   └── enhanced_expressions.json   # O3-enhanced expressions
└── {PATCH_ID}_obj_{OBJ_ID}/
    ├── {PATCH_ID}_bbox.png         # Bounding box view
    ├── {PATCH_ID}_focused.png      # Cropped object view
    └── enhanced_expressions.json   # O3-enhanced expressions

Usage

Used for QLoRA finetuning Gemma3:

python gemma3_lora_finetune.py \
  --enhanced_data_dir enhanced_annotations_o3_dual \
  --output_dir ./gemma-aerial-12b

Cost Efficiency

Generating 500 o3 samples costs ~$10.36. The distilled Gemma3 model processes 259,709 expressions for ~$22.80 (vs $5,382.98 for direct o3 usage), achieving 238× cost reduction.

Citation

@article{marnoto2025aeriald,
  title={Generalized Referring Expression Segmentation on Aerial Photos},
  author={Marnoto, Luís Pedro Soares},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (J-STARS)},
  year={2025},
  note={Submitted}
}