MMGeo: Multimodal Compositional Geo-Localization for UAVs

• Part I: Code Released
• Part II: Dataset

Table of contents

• Dataset Access
• Quick Start
• Citation

Dataset Access

Coming Soon.

Quick Start

cd MMGeo
pip install -r requirements.txt

First initialize the visual model from game4loc training

# GTA-UAV-MM cross-area setting
python train_gta.py \
    --data_root <The directory of the GTA-UAV-MM dataset> \
    --train_pairs_meta_file "cross-area-drone2sate-train.json" \
    --test_pairs_meta_file "cross-area-drone2sate-test.json" \
    --model "vit_base_patch16_rope_reg1_gap_256.sbb_in1k" \
    --gpu_ids 0,1 --lr 0.0001 --batch_size 64 \
    --with_weight --k 5 --epoch 5

Then do the multimodal training with text modality

# with text
python train_gta_mm.py \
    --data_root <The directory of the GTA-UAV-MM dataset> \
    --train_pairs_meta_file "cross-area-drone2sate-train.json" \
    --test_pairs_meta_file "cross-area-drone2sate-test.json" \
    --model "vit_base_patch16_rope_reg1_gap_256.sbb_in1k" \
    --checkpoint_start <pretrained visual model .pth> \
    --with_text --token_length 50 \
    --gpu_ids 0,1 --lr 0.0001 --batch_size 64 \
    --with_weight --k 5 --epoch 5

Or with point cloud by setting --with_pc, with depth by --with_depth.

Citation

Coming Soon.