FloCard Tree Augmentation Solution with GAN Pipeline

[ayushbariyar]

Approach

Used a hybrid approach combining classical augmentation (Pillow)
and DCGAN (PyTorch) to generate 1355+ image variations from 5 seed images.

What is implemented

• 13 augmentation types: brightness, blur, rotation, crop, contrast,
  rain, fog, color temperature, occlusion, season, time of day,
  health conditions, noise, motion blur
• DCGAN trained for 200 epochs — generates brand new tree images
• Full metadata tracking in manifest.jsonl (12 fields per image)
• Visual HTML gallery with filter buttons
• Dataset export as ZIP
• 43 automated tests — all passing

How to run

pip install Pillow torch torchvision
python augment.py
python gallery.py
python gan_train.py
python gan_generate.py
python test_pipeline.py

Assumptions and limitations

• Trained on 5 seed images — more images would improve GAN quality
• Weather overlays are simulated not photorealistic
• CPU training only — GPU would be faster

Data used

• Real seed images provided by FloCard team
• No private geo-tagged data committed