Fork note. This is a fork of justchenhao/BIT_CD, the official implementation of "Remote Sensing Image Change Detection with Transformers" (Chen, Qi & Shi, IEEE TGRS 2021). All credit for the method and original code belongs to the authors. Everything in this section is my own reproduction and extension; the authors' original README follows below.
I reproduced BIT-CD on LEVIR-CD, then asked whether a foundation-model backbone would help it generalize to a different change-detection dataset without retraining.
Reproduction. Trained BIT on LEVIR-CD and matched the published result (F1 0.9027 vs the paper's 0.8931), after three one-line compatibility fixes for PyTorch 2.x / modern numpy. Details in reproduction.md.
The problem. That same LEVIR-trained model, run zero-shot on WHU-CD, drops ~20 F1 points (0.903 → 0.704). Precision falls further than recall, so out-of-domain it over-flags change.
The experiment. I swapped BIT's ImageNet ResNet backbone for two foundation models, frozen and fine-tuned, with a small adapter reshaping ViT patch tokens into BIT's spatial feature map: DOFA (Earth-observation pretrained) and DINOv2 (general-purpose vision).
| Backbone | LEVIR F1 | WHU F1 (zero-shot) | Keeps |
|---|---|---|---|
| ResNet (BIT baseline) | 0.903 | 0.704 | 78% |
| DOFA frozen | 0.721 | 0.693 | 96% |
| DOFA fine-tuned | 0.724 | 0.518 | 72% |
| DINOv2 frozen | 0.773 | 0.781 | 101% |
| DINOv2 fine-tuned | 0.803 | 0.386 | 48% |
The fine-tuned model is the better of the two in-domain and the worse of the two the moment the domain changes — visible on a WHU test scene, where fine-tuned DINOv2 floods the mask with false positives while the frozen model holds its shape:
Finding. My hypothesis was that the EO-specific model would transfer better. It did not. The axis that mattered was frozen vs fine-tuned, not EO vs general: frozen features held up across the domain shift while fine-tuned ones collapsed (fine-tuned DINOv2's precision fell to 0.266, over-flagging unchanged buildings), and general-purpose DINOv2 beat EO-specific DOFA.
Caveats. Single seed, one dataset pair, one transfer direction, and a deliberately simple adapter that caps in-domain F1. This is research-grade evidence, not a benchmark. Full caveats and the failure-mode figures are in bit-cd-foundation-results.md; the code is in foundation_backbone_comparison.ipynb and dofa_bit_integration.ipynb.
Here, we provide the pytorch implementation of the paper: Remote Sensing Image Change Detection with Transformers.
For more ore information, please see our published paper at IEEE TGRS or arxiv.
Python 3.6
pytorch 1.6.0
torchvision 0.7.0
einops 0.3.0
Clone this repo:
git clone https://github.com/justchenhao/BIT_CD.git
cd BIT_CDWe have some samples from the LEVIR-CD dataset in the folder samples for a quick start.
Firstly, you can download our BIT pretrained model——by baidu drive, code: 2lyz or google drive. After downloaded the pretrained model, you can put it in checkpoints/BIT_LEVIR/.
Then, run a demo to get started as follows:
python demo.py After that, you can find the prediction results in samples/predict.
You can find the training script run_cd.sh in the folder scripts. You can run the script file by sh scripts/run_cd.sh in the command environment.
The detailed script file run_cd.sh is as follows:
gpus=0
checkpoint_root=checkpoints
data_name=LEVIR # dataset name
img_size=256
batch_size=8
lr=0.01
max_epochs=200 #training epochs
net_G=base_transformer_pos_s4_dd8 # model name
#base_resnet18
#base_transformer_pos_s4_dd8
#base_transformer_pos_s4_dd8_dedim8
lr_policy=linear
split=train # training txt
split_val=val #validation txt
project_name=CD_${net_G}_${data_name}_b${batch_size}_lr${lr}_${split}_${split_val}_${max_epochs}_${lr_policy}
python main_cd.py --img_size ${img_size} --checkpoint_root ${checkpoint_root} --lr_policy ${lr_policy} --split ${split} --split_val ${split_val} --net_G ${net_G} --gpu_ids ${gpus} --max_epochs ${max_epochs} --project_name ${project_name} --batch_size ${batch_size} --data_name ${data_name} --lr ${lr}You can find the evaluation script eval.sh in the folder scripts. You can run the script file by sh scripts/eval.sh in the command environment.
The detailed script file eval.sh is as follows:
gpus=0
data_name=LEVIR # dataset name
net_G=base_transformer_pos_s4_dd8_dedim8 # model name
split=test # test.txt
project_name=BIT_LEVIR # the name of the subfolder in the checkpoints folder
checkpoint_name=best_ckpt.pt # the name of evaluated model file
python eval_cd.py --split ${split} --net_G ${net_G} --checkpoint_name ${checkpoint_name} --gpu_ids ${gpus} --project_name ${project_name} --data_name ${data_name}"""
Change detection data set with pixel-level binary labels；
├─A
├─B
├─label
└─list
"""
A: images of t1 phase;
B:images of t2 phase;
label: label maps;
list: contains train.txt, val.txt and test.txt, each file records the image names (XXX.png) in the change detection dataset.
LEVIR-CD: https://justchenhao.github.io/LEVIR/
WHU-CD: https://study.rsgis.whu.edu.cn/pages/download/building_dataset.html
Code is released for non-commercial and research purposes only. For commercial purposes, please contact the authors.
If you use this code for your research, please cite our paper:
@Article{chen2021a,
title={Remote Sensing Image Change Detection with Transformers},
author={Hao Chen, Zipeng Qi and Zhenwei Shi},
year={2021},
journal={IEEE Transactions on Geoscience and Remote Sensing},
volume={},
number={},
pages={1-14},
doi={10.1109/TGRS.2021.3095166}
}