中文 | English

Predict the rotation angle of given picture through CNN. This project can be used for rotate-captcha cracking.

Test result:

Three kinds of model are implemented, as shown in the table below.

RotNet is the implementation of d4nst/RotNet over PyTorch. RotNetR is based on RotNet, with RegNet as its backbone and class number of 128. The average prediction error is 15.1818°, obtained by 64 epochs of training (3 hours) on the Google Street View dataset.

The Cross-Domain Test uses Google Street View and Landscape-Dataset for training, and Captcha Pictures from Baidu (thanks to @xiangbei1997) for testing.

The captcha picture used in the demo above comes from RotateCaptchaBreak

• Device supporting CUDA10+ (mem>=4G for training)

• Python>=3.8,<3.13

• PyTorch>=1.11

• Clone the repository.

git clone https://github.com/Starry-OvO/rotate-captcha-crack.git --depth=1
cd ./rotate-captcha-crack

• Install all requiring dependencies.

This project strongly suggest you to use rye for package management. Run the following commands if you already have the rye:

rye pin 3.12
rye sync

Or, if you prefer conda: The following steps will create a virtual env under the working directory. You can also use a named env.

conda create -p .conda
conda activate ./.conda
conda install matplotlib tqdm tomli
conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia

Or, if you prefer a direct pip:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install -e .

Download the *.zip files in Release and unzip them all to the ./models dir.

The directory structure will be like ./models/RotNetR/230228_20_07_25_000/best.pth

The names of models will change frequently as the project is still in beta status. So, if any FileNotFoundError occurs, please try to rollback to the corresponding tag first.

If no GUI is presented, try to change the debugging behavior from showing images to saving them.

rye run python test_captcha.py

If you do not have the rye, just strip the prefix rye run.

• Install extra dependencies

With rye:

rye sync --features=server

or with conda:

conda install aiohttp httpx[cli]

or with pip:

pip install -e .[server]

• Launch server

rye run python server.py

• Another Shell to Send Images

rye run httpx -m POST http://127.0.0.1:4396 -f img ./test.jpg

• For this project I'm using Google Street View and Landscape-Dataset for training. You can collect some photos and leave them in one directory. Without any size or shape requirement.

• Modify the dataset_root variable in train.py, let it points to the directory containing images.

• No manual labeling is required. All the cropping, rotation and resizing will be done soon after the image is loaded.

rye run python train_RotNetR.py

rye run python test_RotNetR.py

Most of the rotate-captcha cracking methods are based on d4nst/RotNet, with ResNet50 as its backbone. RotNet treat the angle prediction as a classification task with 360 classes, then use cross entropy to compute the loss.

Yet CrossEntropyLoss over one-hot labels will bring a uniform metric distance between any angles (e.g. $\mathrm{dist}(1°, 2°) = \mathrm{dist}(1°, 180°)$ ), clearly defies our common sense. Arbitrary-Oriented Object Detection with Circular Smooth Label (ECCV'20) introduces an interesting trick, by smoothing the one-hot label, e.g. [0,1,0,0] -> [0.1,0.8,0.1,0], CSL provides a loss measurement closer to our intuition, such that $\mathrm{dist}(1°,180°) \gt \mathrm{dist}(1°,3°)$.

Meanwhile, the angle_error_regression proposed by d4nst/RotNet is less effective. That's because when dealing with outliers, the gradient leads to a non-convergence result. It's better to use a SmoothL1Loss for regression.