World-UAV-ds/GeoLoc-UAV-main

GeoLoc-UAV (GeoLoc-UAV-main)

Research code for UAV geo-localization / cross-view image retrieval on the UAV-GeoLoc (World-UAV) dataset.

• Dataset: available on Hugging Face: https://huggingface.co/datasets/RingoWRW97/UAV-GeoLoc
• This folder (GeoLoc-UAV-main) contains training/evaluation scripts and model definitions.

What this code does

The task is formulated as retrieval: given a UAV query image, retrieve the matching geo-referenced database (DB) image(s).

• Training: contrastive classification via InfoNCE (implemented as CrossEntropyLoss on similarity matrix).
• Evaluation: extract global descriptors for queries and DB, then run FAISS IndexFlatL2 search and report Top-1 / Top-5 / Top-10 accuracy.

Two main modes exist in the code:

• vanilia (spelling in code): a standard CNN/ViT backbone (resnet18, dinov2_*) + an aggregation head (multiconvap, convap, optional LPN).
• group: a GroupNet-style encoder that uses a set of transformed views + point grids (scale/rotate sampling) before aggregation.

Project structure

Key entry points:

• train_vanilia.py: train a vanilla backbone (ResNet) retrieval model.
• train_vanilia_dino.py: same, but with a DINOv2 backbone.
• train_group.py: train the group-based model.
• train_group_dino.py: group-based model with DINOv2 features.
• preprocess_data.py: helper to generate train index files from scene lists.

Evaluation:

• eval_simidataset_parser.py: evaluate on World-UAV-style splits (reads a list of scene folders).
• eval_real_dataset.py: evaluate on a “real” dataset layout (query_images/reference_images + gt CSV/NPY).
• eval_denseuav.py: evaluate on DenseUAV-style lists (query.txt/db.txt/gt.txt).
• eval_real.sh, eval_rot.sh: example command lines (paths are author-specific).

Core modules:

• dataset/World.py: dataset loaders for World-UAV, “real” and DenseUAV layouts.
• models/: backbones, aggregators, group networks.
• eval/eval.py: feature extraction + FAISS retrieval metrics.

Installation

This repository does not ship a pinned requirements.txt. A typical working environment:

python -m venv .venv
source .venv/bin/activate

pip install -U pip
pip install torch torchvision transformers tensorboard tqdm pillow numpy pandas h5py matplotlib opencv-python

# FAISS (choose one):
pip install faiss-cpu
# or: pip install faiss-gpu  (if your platform provides it)

Notes:

• Most scripts assume CUDA is available; CPU should work for evaluation but may be slow.
• Mixed precision is enabled by default in training scripts.

Dataset format (World-UAV style)

For the World-UAV dataset loaders (WorldDatasetEvalVanilia / WorldDatasetEvalGroup), the code expects a scene folder with:

<dataset_root_dir>/
  <scene_name>/
    positive.json
    semi_positive.json
    DB/
      img/
        *.png
    query/
      height100_rot0/footage/*.jpeg
      height100_rot45/footage/*.jpeg
      ...
      height150_rot315/footage/*.jpeg

Ground truth:

• positive.json: maps each query key to a list of positive DB image filenames (under DB/img/).
• semi_positive.json: optional extra positives (the code includes them in GT if present).

Important:

• Query images are collected across multiple height/rotation folders. The evaluator replicates GT to match the number of query images extracted (see eval/eval.py: multi_num = ql.shape[0] / len(pos_gt)).

Training index files (World-UAV style)

Training loaders (WorldDatasetTrainVanilia, WorldDatasetTrainGroup) read a text file where each line contains:

<relative_query_path> <label_int> <relative_db_path>

Example (paths are relative to dataset_root_dir):

some_scene/query/height100_rot0/footage/height100_rot0_0001.jpeg 12 some_scene/DB/img/000123.png

The helper script preprocess_data.py can generate *_query_all.txt and *_db_all.txt from a file containing scene names (one per line). It is currently written with author-specific absolute paths and may require edits to:

• root
• txt (scene list file)
• save_path

Quick start (training)

All training scripts contain a Configuration dataclass with hardcoded paths like:

• dataset_root_dir
• train_query_txt
• val_index_txt
• test_index_txt

Update them to your local paths before running.

Vanilla (ResNet)

python train_vanilia.py

Vanilla (DINOv2 backbone)

python train_vanilia_dino.py

Group model

python train_group.py

Group model (DINO variant)

python train_group_dino.py

Checkpoints:

• Training scripts create a timestamped directory under config.model_path/config.model/<HHMMSS>/
• The best checkpoint (by average Top-1 on validation scenes) is saved as weights_e{epoch}_{score}.pth

TensorBoard:

• Vanilla scripts write under world_vanillia/...
• Group script writes under world/...

Quick start (evaluation)

Evaluate on World-UAV scenes list

Use eval_simidataset_parser.py. It reads a text file where each line is a scene folder name (relative to --dataset_root):

python eval_simidataset_parser.py \
  --mode vanilia \
  --dataset_root "/path/to/WorldLoc" \
  --test_txt "/path/to/WorldLoc/Index/test.txt" \
  --save_txt "/tmp/results.txt" \
  --checkpoint_path "/path/to/weights.pth" \
  --backbone_arch resnet18 \
  --pretrain_flag False \
  --agg_in_channels 512 \
  --agg_out_channels 512 \
  --agg_LPN False

For the group model:

python eval_simidataset_parser.py \
  --mode group \
  --dataset_root "/path/to/WorldLoc" \
  --test_txt "/path/to/WorldLoc/Index/test.txt" \
  --save_txt "/tmp/results.txt" \
  --checkpoint_path "/path/to/weights.pth" \
  --agg_in_channels 256 \
  --agg_out_channels 256

Evaluate on DenseUAV lists

python eval_denseuav.py \
  --mode vanilia \
  --dataset_query "/path/to/query.txt" \
  --dataset_db "/path/to/db.txt" \
  --dataset_gt "/path/to/gt.txt" \
  --checkpoint_path "/path/to/weights.pth"

Evaluate on “real” query/reference folder layout

python eval_real_dataset.py \
  --mode vanilia \
  --dataset_root_dir "/path/to/test_set" \
  --checkpoint_path "/path/to/weights.pth" \
  --save_dir_path "/tmp/geoloc-uav-real-eval"

Common pitfalls / required path fixes

• Hardcoded absolute paths: many scripts use /media/... paths. Replace them with your local paths.
• Transform config path is hardcoded in code:
  • dataset/World.py and models/group/groupnet_dino.py load transform_config.json via an absolute json_path.
  • For portability, point it to configs/transform_config.json in this repository.
• Mode spelling: the scripts use vanilia (not vanilla).

Demo