AdaptOVCD

Training-Free Open-Vocabulary Change Detection via Adaptive Foundation Model Synergy

Paper | arXiv

Overview

AdaptOVCD is a training-free framework for open-vocabulary change detection in remote sensing imagery. By synergistically integrating three foundation models with adaptive enhancement modules, AdaptOVCD enables zero-shot detection of arbitrary change categories specified via natural language—without any task-specific training or annotated data.

Framework Overview

The pipeline implements dual-dimensional multi-level information fusion: vertically constructing a data-feature-decision cascade, and horizontally conducting targeted adaptive designs at each level.

Visualization of Intermediate Processes

To gain a deeper understanding of the mechanism of each module, we visualize the intermediate processes of AdaptOVCD. The visualization demonstrates the progressive filtering effect, where spurious detections are correctly eliminated while authentic changes are retained.

Installation

See INSTALL.md for detailed setup instructions.

Project Structure

AdaptOVCD/
├── models/                                # See INSTALL.md for setup
├── third_party/                           # See INSTALL.md for setup
├── data/                                  # Download required
│   ├── levircd/                           # Ready to use
│   ├── Building change detection dataset_add/  # Preprocessing required
│   ├── DSIFN/                             # Preprocessing required
│   └── second_dataset/                    # Preprocessing required
├── configs/
│   ├── models/                            # Model configurations
│   └── datasets/                          # Dataset configurations
├── utils/
│   └── datasets-test/                     # Preprocessing scripts
│       ├── dsifn.py
│       ├── second.py
│       └── whucd.py
├── evaluate.py                            # LEVIR-CD, WHU-CD, DSIFN evaluation
├── evaluate_second.py                     # SECOND evaluation
└── demo.py                                # Demo script

Datasets

We use test sets only for zero-shot evaluation. Download and organize as follows:

Dataset	Type	Test Size	Resolution	Download
LEVIR-CD	Building	128 pairs	0.5m	Link
WHU-CD	Building	690 pairs	0.075m	Link
DSIFN	Building	48 pairs	2m	Link
SECOND	Semantic (6 classes)	1000+ pairs	-	Link

Raw Dataset Structure

After downloading, place datasets in data/ directory:

data/
├── levircd/                          # LEVIR-CD (ready to use)
│   ├── A/
│   ├── B/
│   └── label/
├── Building change detection dataset_add/   # WHU-CD (needs preprocessing)
│   └── 1. The two-period image data/
│       ├── 2012/splited_images/test/image/
│       ├── 2016/splited_images/test/image/
│       └── change_label/test/
├── DSIFN/                            # DSIFN (needs preprocessing)
│   └── test/
│       ├── t1/
│       ├── t2/
│       └── mask/
└── second_dataset/                   # SECOND (needs preprocessing)
    └── test/
        ├── im1/
        ├── im2/
        ├── label1/
        └── label2/

Dataset Preprocessing

Run preprocessing scripts to convert datasets into unified format (A/B/label):

# WHU-CD: Convert TIF to PNG and reorganize structure
python utils/datasets-test/whucd.py --source data/Building\ change\ detection\ dataset_add --output data/whucd

# DSIFN: Convert mask TIF to JPG and reorganize
python utils/datasets-test/dsifn.py --source data/DSIFN/test --target data/dsifn

# SECOND: Generate class-specific change labels from semantic labels
python utils/datasets-test/second.py
# Input: data/second_dataset/test -> Output: data/second/

Preprocessed Dataset Structure

After preprocessing:

data/
├── levircd/          # Ready to use
│   ├── A/
│   ├── B/
│   └── label/
├── whucd/            # After whucd.py
│   ├── A/
│   ├── B/
│   └── label/
├── dsifn/            # After dsifn.py
│   ├── A/
│   ├── B/
│   └── label/
└── second/           # After second.py
    ├── im1/
    ├── im2/
    ├── label1/
    ├── label2/
    ├── label_building/
    ├── label_water/
    ├── label_tree/
    ├── label_low_vegetation/
    ├── label_non_veg_ground_surface/
    └── label_playground/

Usage

Demo

python demo.py \
    --model OVCD_levircd \
    --input1 demo_images/A/00004.png \
    --input2 demo_images/B/00004.png \
    --output outputs/demo/

Evaluation

Binary change detection (LEVIR-CD, WHU-CD, DSIFN):

# LEVIR-CD
python evaluate.py --model OVCD_levircd --dataset levircd

# WHU-CD
python evaluate.py --model OVCD_whucd --dataset whucd

# DSIFN
python evaluate.py --model OVCD_dsifn --dataset dsifn

Semantic change detection (SECOND - 6 classes):

# Single class
python evaluate_second.py --class building --output_dir outputs/second/building

# All 6 classes
python evaluate_second.py --class all --output_dir outputs/second/all

Available SECOND classes: building, water, tree, low_vegetation, non_veg_ground_surface, playground

Qualitative Results

Building Change Detection on LEVIR-CD, WHU-CD, DSIFN, and SECOND (Building):

Open-Vocabulary Change Detection on SECOND across six semantic categories (Building, Low Vegetation, Non-veg. Ground, Playground, Tree, Water):

Save Predictions

Add --save_predictions flag to keep output masks:

python evaluate.py --model OVCD_levircd --dataset levircd --save_predictions --output_dir outputs/

Citation

@article{adaptovcd2025,
  title={AdaptOVCD: Training-Free Open-Vocabulary Change Detection via Adaptive Foundation Model Synergy},
  author={},
  journal={},
  year={2025}
}

Acknowledgements

• SAM-HQ (Apache 2.0)
• DINOv3 (Meta License)
• DGTRS-CLIP (Apache 2.0)
• CLIP (MIT)
• DynamicEarth

License

Apache 2.0