Change Detection Laboratory

Yet another project for developing and benchmarking deep learning-based remote sensing change detection methods.

CDLab also has a PaddlePaddle version.

News

Add official implementation of P2V-CD (TIP 2022).

Prerequisites

opencv-python==4.1.1
pytorch==1.6.0
torchvision==0.7.0
pyyaml==5.1.2
scikit-image==0.15.0
scikit-learn==0.21.3
scipy==1.3.1
tqdm==4.35.0

Tested using Python 3.7.4 on Ubuntu 16.04.

Get Started

In src/constants.py, change the dataset locations to your own.

Data Preprocessing

In scripts/ there are preprocessing scripts for several datasets。

Model Training

To train a model from scratch, use

python train.py train --exp_config PATH_TO_CONFIG_FILE

A few configuration files regarding different datasets and models are provided in the configs/ folder for ease of use. Note that the hyperparameters are not elaborately investigated to obtain a fully optimized performance.

As soon as the program starts and prints out the configurations, there will be a prompt asking you to write some notes. What you write will be recorded into the log file to help you remember what you did, or you can simply skip this step by pressing [Enter].

To resume training from some checkpoint, run the code with the --resume option.

python train.py train --exp_config PATH_TO_CONFIG_FILE --resume PATH_TO_CHECKPOINT

Other frequently used commandline options include:

--anew: Add it if the checkpoint is just used to initialize model weights. Note that loading an incompatible checkpoint is supported as a feature, which is useful when you are trying to utilize a well pretrained model for finetuning.
--save_on: By default, an epoch-based trainer is used for training. At the end of each training epoch, the trainer evaluates the model on the validation dataset. If you want to save the model output during the evaluation process, enable this option.
--log_off: Disable logging.
--tb_on: Enable TensorBoard summaries.
--debug_on: Useful when you are debugging your own code. In debugging mode, no checkpoint or model output will be written to disk. In addition, a breakpoint will be set where an unhandled exception occurs, which allows you to locate the causes of the crash or do some cleanup jobs.

During or after the training process, you can check the model weight files in exp/DATASET_NAME/weights/, the log files in exp/DATASET_NAME/logs/, and the output change maps in exp/DATASET_NAME/out/.

Model Evaluation

To evaluate a model on the test subset, use

python train.py eval --exp_config PATH_TO_CONFIG_FILE --resume PATH_TO_CHECKPOINT --save_on --subset test

This project also provides the funtionality of sliding-window test on large raster images. Use the following command:

python sw_test.py --exp_config PATH_TO_CONFIG_FILE \
  --resume PATH_TO_CHECKPOINT --ckp_path PATH_TO_CHECKPOINT \
  --t1_dir PATH_TO_T1_DIR --t2_dir PATH_TO_T2_DIR --gt_dir PATH_TO_GT_DIR

Other frequently used commandline options of src/sw_test.py include:

--window_size: Set the size of the sliding window.
--stride: Set the stride of the sliding window.
--glob: Specify a wildcard pattern to match files in t1_dir, t2_dir, and gt_dir.
--threshold: Set the threshold used to convert the probability map to the change map.

Note however that currently src/sw_test.py does not support custom pre-processing or post-processing modules.

Use Models from Third-Party Libraries

Currently this projects supports the training and evaluation of models from the change_detection.pytorch library, which can be achieved by simply modifying the configuration files. Please refer to the example in configs/svcd/config_svcd_cdp_unet.yaml。

The version number of the supported change_detection.pytorch library is 0.1.0.

Supported Models

Model	Name	Link
CDNet	`CDNet`	paper
FC-EF	`UNet`	paper
FC-Siam-conc	`SiamUNet-conc`	paper
FC-Siam-diff	`SiamUNet-diff`	paper
STANet	`STANet`	paper
DSIFN	`IFN`	paper
SNUNet	`SNUNet`	paper
BIT	`BIT`	paper
L-UNet	`LUNet`	paper
DSAMNet	`DSAMNet`	paper
P2V-CD	`P2V`	paper

Supported Datasets

Dataset	Name	Link
Synthetic images and real season-varying remote sensing images	`SVCD`	google drive
WHU building change detection dataset	`WHU`	website
LEVIR building change detection dataset	`LEVIRCD`	website

Performance on SVCD Dataset

Model	Precision	Recall	F1	OA
CDNet	92.99	87.08	89.94	97.59
FC-EF	94.28	83.80	88.73	97.37
FC-Siam-conc	94.57	91.34	92.93	98.28
FC-Siam-diff	95.87	90.60	93.16	98.36
STANet	89.22	98.25	93.52	98.32
DSIFN	97.64	96.35	96.99	99.26
SNUNet	97.89	97.25	97.57	99.40
BIT	97.20	96.38	96.79	99.21
L-UNet	96.48	94.79	95.63	98.93
DSAMNet	92.78	98.06	95.35	98.82
P2V-CD	98.57	98.26	98.42

License

This project is released under the Unlicense.

Contributing

Any kind of contributions to improve this repository is welcome.

Bobholamovic/CDLab