A python based annotation/labelling toolbox for images. The program allows the user to annotate individual objects in images.
Currently the pychet labeller supports Circle Annotation and Rectangle Annotation.
The circle annotation enables the user to annotate objects in an image with its centroid and its radius. This makes the annotation toolbox particularly suitable for round objects such as fruits, balls, coins etc.
The rectangle annotation annotates a box with top left position and its width and height. This is more standard to what is used in the ML community where a bounding box is scribed around objects.
There are other examples out there (which also use PyQt) for labelling images. Generally I found that they offered a lot of flexibility in the types of annotations that could be done. However, this also meant that labelling very basic things was very slow. For example, many tools allow you to annotate the vertices of polygons and then type in the object name. However, the dataset I created this for contained circular objects. All I wanted was to annotate a centroid and object radius. For the rectangles, I didn't want to click on the vertices, which can be difficult for complex shaped objects.
Additionally, I wanted a toolbox where you could see your bounding box/bounding region live while moving the mouse around. And wanted some simple shortcuts to changes its size and change the labelling output. Pychet Labeller aims to address these features.
sudo apt-get install libgeos-dev libgeos++-dev python-pip python2.7-dev libxext-dev python-qt4 qt4-dev-tools build-essential
sudo -H pip install -U svgwrite shapely simplejson
- Install prerequisites (svgwrite)
- Clone this repository git clone https://github.com/acfr/pychetlabeller.git pychetlabeller
- Build and install python setup.py build && sudo python setup.py install
python -m pychetlabeller <img dir> <label dir> --tool <circle | rectangle> --labelmap <labelmap.json>
See src/pychetlabeller/sample_labelmap.json for example labelmap file
Pychet Labeller makes it very easy to label a group of images in a folder, one after the other. Simply run the labeller and open up the images directory form the push button on the top right.
Annotations are made by simple clicking on the image. The user has the option to move the image, change the size of the annotation tool, zoom in/out of the image, change the object label, save the label or go to the next image. Press F1 to view the shortcuts for these things.
For circles, the size is the radius, and for rectangles, the size is the width and height.
A few notes:
- Currently the program will automatically detect any files with extensions: png, jpg, , jpeg, tiff, bmp
- When labelling multiple images, can enable save_label to automatically save the labels - otherwise press ctrl-x to save current annotations
- Individual annotations can be deleted by selecting them on the table (or shift clicking on the image) and pressing delete.
- Backspace deletes the last annotation added
The annotations are saved in csv format with the same name as the input image file. The csv entries for circles are item, centre-x, centre-y, radius, label id. The csv entries for rectangles are item, topleft-x, topleft-y, width, height, label id. The annotations are also automatically saved in .svg format
By default the annotations are saved in the image parent directory under a new folder: labels. The user can choose to manually set a different folder for the labels.
We can also edit objects/object_labeler.py to quickly label one image. Under the object MainWindow, uncomment self.quickview(), then under function quickview() set your image path.
Extentions to labeller - coming soon:
- Currently can choose only one or the other - circles or rectangles - due to strict csv format. Should resort to svg only format and save all shapes
- Add generic polygon shape - and test other shapes
- Change brightness and contrast sliders to levels slider. Allows for much better control of the image contrast
Please contact author to report bugs @ bargoti.suchet@gmail.com