Image auto crop toolbox. Supports image cropping with any aspect ratio, based on face-detection and GAIC: Grid-Anchor-based-Image-Cropping-Pytorch.

This project only supports python3 and pytorch > 1.0

• Setup
• Demo
• Different from GAIC
• Reference
• Citation

You can choose to install via pypi or install it via source code.

auto-crop can be installed directly through pypi

pip install auto-crop

# colone repository
git clone https://github.com/lih627/autocrop.git
# install autocrop
cd /path/to/autocrop
python setup.py install

Note : If there is errors when compiling CPP and CUDA extensions, you can choose to compile CPP/CUDA api separately.

cd autocrop/model/rod_align
python setup.py install
cd ../roi_aligm
python setup.py install

Here is a simple demo. From demo.py

First, build a cropper. Cropper contains GAIC pretrained models, and you can select load a DSFD Face-Detecor or not.

from autocrop import cropper
autocropper = cropper.AutoCropper(model='mobilenetv2', # 'mobilenetv2' or 'shufflenetv2'
                                  cuda=True, # if GUDA is avaliable and True, Inference on GPU
                                  use_face_detector=True) # Use Face Detector to filter RoIs

Then, use cropper to crop RGB formate image. The selectable parameters are the number of cropping results, the aspect ration and whether to used face detection results to assist in generating RoIs. crop_ret is a list with sizetopK x 4 , Each cropping result is encoded as [xmin, ymin, xmax, ymax] in pixel coordinate system.

import cv2
# BGR to RGB
img = cv2.imread('imgs/demo.jpg')
img_ = img[:, :, (2, 1, 0)]
# get crop result
crop_ret = autocropper.crop(img_,
                            topK=1,
                            crop_height=1,
                            crop_width=1,
                            filter_face=True, # True: Crop result will not contain half face
                            single_face_center=True) # True: face in the crop result's width center

You can visualize the cropping results

for bbox in crop_ret:
    r, g, b = int(random.random() * 255), int(random.random() * 255), int(random.random() * 255)
    cv2.rectangle(img, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (b, g, r))
cv2.imshow('ret', img)
cv2.waitKey()

This project is mostly based on GAIC, and the modules are listed as follows:

It is slightly different from GAIC in practice, as shown below:

1. We can specify any crop ratio

GAIC supports RoIs with uncertain aspect ratios and several RoIs with fixed aspect ratios(1:1, 4:3, 16:9). In practical applications, image cropping needs to select the cropping area according to the fixed aspect ratio. I modified the code of the bboxes generation part. For RoIs evaluation, I used the GAIC pre-trained model.

2. If there is only half a face in the bounding box, filter out the bounding box

At the same time, in practical applications, when the distribution of people in the picture is not fixed, for example, when two people stand on the left and right sides of the picture, the RoI selected by GAIC may tear the human body. We adopt the face detection method to filter out some non-conformities. The required RoI will be evaluated after.

There is a comparison:

3. When the bounding box has only one face, the face should be in the middle of the box as much as possible

We have added additional options when generating anchor boxes. If there is only one face in a RoI, use the RoI with the face in the middle of the RoI's width direction. see autocrop/cropper.py for details.

There is a comparison:

1. GAIC: Grid-Anchor-based-Image-Cropping-Pytorch MIT License
2. DSFD-Pytorch-Inference Apache-2.0 License

If you find this code useful, remember to cite the original authors:

For GAIC:

@inproceedings{zhang2019deep,
  title={Reliable and Efficient Image Cropping: A Grid Anchor based Approach},
  author={Zeng, Hui, Li, Lida， Cao, Zisheng and Zhang, Lei},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

For DSFD:

@inproceedings{li2018dsfd,
  title={DSFD: Dual Shot Face Detector},
  author={Li, Jian and Wang, Yabiao and Wang, Changan and Tai, Ying and Qian, Jianjun and Yang, Jian and Wang, Chengjie and Li, Jilin and Huang, Feiyue},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}