zuordnung-data

Zuordnung — Siamese DINOv2 Satellite/Aerial Alignment Models

Pretrained models for matching aerial drone photos to satellite imagery. Each model embeds aerial and satellite images into a shared 256-dimensional space; at inference time, the closest satellite tile to a drone photo is retrieved.

Architecture

Models

Ensemble Configuration

The default ensemble pairs the two strongest complementary models:

[
  {
    "id": "336_gem_v9d_lora_r64",
    "weight": 1.0,
    "role": "primary"
  },
  {
    "id": "336_gem_v4",
    "weight": 1.0,
    "role": "secondary"
  }
]

from gui_app.config import ENSEMBLE_MODELS  # already wired

Loading a Model

import torch, sys
sys.path.insert(0, '..')  # project root
from core.model import DINOv2SiameseNetwork, apply_lora
from core.checkpoint import inspect_checkpoint

path = '336_gem_v9d_lora_r64/336_gem_v9d_lora_r64_best.pth'
info = inspect_checkpoint(path)
model = DINOv2SiameseNetwork(
    backbone=info['backbone'],
    img_size=info['img_size'],
    embedding_dim=info['embedding_dim'],
    aerial_input_dim=info['aerial_input_dim'],
    pool=info['pool'],
    dropout=0.0,
    grad_checkpointing=False,
)

# Apply LoRA if the checkpoint contains LoRA weights
ckpt = torch.load(path, map_location='cpu', weights_only=False)
if any('.attn.qkv.lora_A' in k for k in ckpt['model_state_dict']):
    args = ckpt.get('args', {})
    apply_lora(model,
               rank=args.get('lora_rank', 64),
               alpha=args.get('lora_alpha', args.get('lora_rank', 64)),
               target=args.get('lora_target', 'qv'))

model.load_state_dict(ckpt['model_state_dict'])
model.eval()

Data Splits

• Train: 19,675 pairs (33 sessions)
• Val: 943 pairs (full-gallery, session-split)
• Test: 482 pairs (session-split, no leakage)

Citation

If you use these models, please cite the Zuordnung project.

License

Models are released under the same license as this repository.