DINOMamba / FMB-CD

Foundation Model Bridged Change Detection via Semantic-Guided Detail Injection and Cross-Stream Feature Aggregation.

This repository contains the binary remote-sensing change detection implementation used for our FMB-CD paper draft. The code is adapted from ChangeMamba and replaces the fully fine-tuned VMamba encoder with a parameter-efficient DINOv2-based dual-stream encoder.

Overview

Remote-sensing change detection needs both high-level semantic robustness and fine spatial details. Full fine-tuning of a large backbone can be accurate, but it is expensive. Freezing a foundation model is efficient, but patch-based ViT features may blur boundaries. A CNN detail branch can recover high-frequency cues, but naive fusion may amplify background noise.

FMB-CD addresses this trade-off with:

• Frozen DINOv2 semantic stream with LoRA adapters for lightweight domain adaptation.
• Trainable CNN detail stream for local boundaries and high-frequency spatial cues.
• Semantic-guided detail injection (SGDI) to suppress task-irrelevant detail responses before fusion.
• Cross-stream feature aggregation (CSFA) to align and recalibrate heterogeneous ViT/CNN features.
• Mamba-based decoder inherited from ChangeMamba for multi-scale feature aggregation.

In the current code, the binary change detection path is implemented mainly in:

• changedetection/models/ChangeMambaBCD.py
• changedetection/models/DINO_backbone.py
• changedetection/models/ChangeDecoder.py

The SCD/BDA scripts and VMamba-related modules are retained from the original ChangeMamba codebase for reference, but the README below focuses on the FMB-CD binary change detection setup.

Paper Results

The following numbers are from the current FMB-CD manuscript.

Dataset	OA	F1	IoU	Recall	Precision
LEVIR-CD+	98.90	85.90	75.29	83.63	88.30
WHU-CD	99.49	92.76	86.50	92.14	93.38
SYSU-CD	92.85	84.32	72.89	81.55	87.29

Efficiency comparison:

Method	Trainable Params	Training Iters
ChangeMamba-Base	84.7M	240k
FMB-CD	17.55M	40k

Note: FMB-CD contains a frozen DINOv2 backbone. Trainable parameters are reported because they better reflect the optimization cost.

Repository Layout

DINOMamba/
+-- changedetection/
|   +-- configs/vssm1/              # Mamba decoder configs
|   +-- datasets/                   # BCD/SCD/BDA dataloaders
|   +-- models/
|   |   +-- ChangeMambaBCD.py       # FMB-CD BCD model entry
|   |   +-- DINO_backbone.py        # DINOv2 + LoRA + detail branch
|   |   +-- ChangeDecoder.py        # Mamba decoder
|   +-- script/
|       +-- train_MambaBCD.py       # BCD training
|       +-- infer_MambaBCD.py       # BCD inference
+-- classification/                 # VMamba backbone code inherited from ChangeMamba
+-- kernels/selective_scan/         # CUDA selective scan kernels
+-- pretrained_weight/              # DINOv2/checkpoints
+-- data/                           # optional local dataset lists/data
+-- requirements.txt

Installation

The selective scan kernel is designed for Linux + CUDA. Windows can be used for editing and lightweight inspection, but training is expected to run on a CUDA-enabled Linux environment.

conda create -n dinomamba python=3.8 -y
conda activate dinomamba

# Install PyTorch according to your CUDA version first.
# Example only:
# pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt
pip install peft "timm>=0.9.0"

cd kernels/selective_scan
pip install .
cd ../..

Important dependency note:

• The original ChangeMamba dependency list pins timm==0.4.12.
• FMB-CD uses timm.create_model("vit_base_patch14_dinov2"), so a DINOv2-capable timm version is required.
• peft is required for LoRA injection.

Pretrained Weights

Place the DINOv2 ViT-B/14 checkpoint at:

pretrained_weight/dinov2_vitb14_pretrain.pth

The current implementation loads this path in changedetection/models/DINO_backbone.py.

The old ChangeMamba VMamba checkpoints are not used by the FMB-CD BCD encoder. Use --resume only for checkpoints trained with this DINOMamba/FMB-CD model.

Data Preparation

For binary change detection, organize SYSU-CD, LEVIR-CD+, or WHU-CD as:

DATASET_ROOT/
+-- train/
|   +-- T1/
|   |   +-- 00001.png
|   |   +-- ...
|   +-- T2/
|   |   +-- 00001.png
|   |   +-- ...
|   +-- GT/
|       +-- 00001.png
|       +-- ...
+-- test/
|   +-- T1/
|   +-- T2/
|   +-- GT/
+-- train_list.txt
+-- test_list.txt

The GT mask should use 0 for unchanged pixels and 255 for changed pixels.

Training

Run from the project root. This is important because the DINOv2 checkpoint path is currently relative to the project root.

Example: train FMB-CD on LEVIR-CD+ for 40k iterations.

python changedetection/script/train_MambaBCD.py \
  --dataset LEVIR-CD+ \
  --batch_size 16 \
  --crop_size 256 \
  --max_iters 40000 \
  --val_interval 1000 \
  --model_type FMB-CD \
  --model_param_path changedetection/saved_models \
  --train_dataset_path data/LEVIR-CD+/train \
  --train_data_list_path data/LEVIR-CD+/train_list.txt \
  --test_dataset_path data/LEVIR-CD+/test \
  --test_data_list_path data/LEVIR-CD+/test_list.txt \
  --cfg changedetection/configs/vssm1/vssm_tiny_224_0229flex.yaml

Example for SYSU-CD:

python changedetection/script/train_MambaBCD.py \
  --dataset SYSU \
  --batch_size 16 \
  --crop_size 256 \
  --max_iters 40000 \
  --val_interval 1000 \
  --model_type FMB-CD \
  --model_param_path changedetection/saved_models \
  --train_dataset_path /path/to/SYSU/train \
  --train_data_list_path /path/to/SYSU/train_list.txt \
  --test_dataset_path /path/to/SYSU/test \
  --test_data_list_path /path/to/SYSU/test_list.txt \
  --cfg changedetection/configs/vssm1/vssm_tiny_224_0229flex.yaml

The training script saves:

• best_model.pth
• last_model.pth
• optional visualization outputs under the model save directory

Inference

Use --resume to load an FMB-CD checkpoint.

python changedetection/script/infer_MambaBCD.py \
  --dataset LEVIR-CD+ \
  --model_type FMB-CD \
  --test_dataset_path data/LEVIR-CD+/test \
  --test_data_list_path data/LEVIR-CD+/test_list.txt \
  --cfg changedetection/configs/vssm1/vssm_tiny_224_0229flex.yaml \
  --resume changedetection/saved_models/LEVIR-CD+/FMB-CD_xxx/best_model.pth \
  --result_saved_path results

Predicted binary change maps will be saved to:

results/<dataset>/<model_type>/change_map/

Ablation Protocol

The paper uses the following compact ablation path:

Variant	Recall	Precision	OA	F1	IoU
Baseline (Frozen DINO)	76.60	88.85	98.65	82.27	69.88
+ Detail Stream	85.31	87.67	98.85	85.54	74.73
+ SGDI & CSFA (Full)	83.63	88.30	98.90	85.90	75.29

The intended reading is:

• Frozen DINOv2 alone is semantically robust but misses fine details.
• Adding the detail branch recovers recall.
• The interaction design suppresses false alarms and improves precision.

Common Issues

Issue	Fix
vit_base_patch14_dinov2 is unknown	Install a newer DINOv2-capable timm version.
ModuleNotFoundError: peft	Install peft.
selective_scan_cuda_oflex is not found	Rebuild kernels/selective_scan with the active CUDA/PyTorch environment.
Low inference accuracy	Use --resume with a trained FMB-CD checkpoint. Do not load old VMamba/ImageNet weights as a trained CD model.
CUDA out of memory	Reduce --batch_size, crop size, or disable extra visualization.
DINOv2 checkpoint not found	Put dinov2_vitb14_pretrain.pth under pretrained_weight/ or update the path in DINO_backbone.py.

Citation

The manuscript BibTeX will be added after publication. For now, please cite the repository as:

@misc{yu2026fmbcd,
  title  = {FMB-CD: Foundation Model Bridged Change Detection via Semantic-Guided Detail Injection and Cross-Stream Feature Aggregation},
  author = {Yue Yu and Maoteng Zheng},
  year   = {2026},
  note   = {Manuscript in preparation}
}

This codebase is adapted from ChangeMamba:

@article{chen2024changemamba,
  author  = {Hongruixuan Chen and Jian Song and Chengxi Han and Junshi Xia and Naoto Yokoya},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  title   = {ChangeMamba: Remote Sensing Change Detection with Spatiotemporal State Space Model},
  year    = {2024},
  volume  = {62},
  pages   = {1-20},
  doi     = {10.1109/TGRS.2024.3417253}
}

Acknowledgments

This repository builds on the excellent open-source implementations of ChangeMamba, VMamba, timm, PEFT/LoRA, and DINOv2. Please also follow the licenses of the original projects and datasets.