Automatically annotate your entire image directory by a single command.
For a more detailed explaination and code usage, please refer to this medium article.
As simple as saying - "Annotate all the street sign (label) in the autonomous car dataset (directory)" and BAM! DONE. Each and every image with a street sign in the diverse dataset directory containing images of all sorts which have a street sign are filtered and the segmentation annotation is performed in a single command.
The Auto-Annotate tool provides auto annotation of segmentation masks for the objects in the images inside some directory based on the labels. Auto-Annotate is able to provide automated annotations for the labels defined in the COCO Dataset and also supports Custom Labels. This tool is built on top of Mask R-CNN to support auto annotation for each instance of an object segment in the image.
The Auto-Annotate tool works on two modes
- COCO Label Annotation - NO TRAINING REQUIRED. Just use the weights of the the Coco dataset. Point to the directory and the annotations are ready.
- Custom Label Annotation - Train the model for the custom label. Use the weights and annotate.
NOTE: Please refer to knownIssues.md file in the repo for known issues and their resolution. Please feel free to contribute in case of any errors/issues arising during the installation and usage of the tool.
[
{
"filename": "bird_house_in_lawn.jpg",
"id": 1,
"label": "Bird House",
"bbox": [ 111.5, 122.5, 73, 66 ],
"segmentation": [
[ 167, 188.5, 174, 185.5, 177.5, 181, 177.5, 157, 183.5, 154, 184.5, 149, 159, 124.5, 150, 122.5,
142, 124.5, 131, 131.5, 111.5, 150, 111.5, 156, 116.5, 162, 116.5, 184, 121, 188.5, 167, 188.5
]
]
}
][
{
"filename": image_file_name,
"id": id_of_the_image,
"label": label_to_search_and_annotate,
"bbox": [ x, y, w, h ], -- x,y coordinate of top left point of bounding box
-- w,h width and height of the bounding box
Format correspond to coco json response format for bounding box
"segmentation": [
[ x1, y1, x2, y2,...] -- For X,Y belonging to the pixel location of the mask segment
-- Format correspond to coco json response format for segmentation
]
}
]
| ORIGINAL IMAGE | MASKED IMAGE |
|---|---|
 |
 |
-
Clone this repository
-
Install dependencies
pip3 install -r requirements.txt
-
If annotating objects supported by COCO Dataset Download pre-trained COCO weights (mask_rcnn_coco.h5) from the releases page and store it in the root directory. If annotating objects Custom Objects Train Mask RCNN and use those weights.
-
Run Commands as below based on the mode.
python3 annotate.py annotateCoco --image_directory=/path/to/the/image/directory/ --label=object_label_to_annotate --weights=/path/to/weights.h5 --displayMaskedImages=Falsepython3 annotate.py annotateCustom --image_directory=/path/to/the/image/directory/ --label=object_label_to_annotate --weights=/path/to/weights.h5 --displayMaskedImages=False- Find the annotations in the directory - /path/to/the/image/directory/ specified above
Use pre-trained weights for MS COCO. After finishing with the installation, we can run it directly from the command line as such:
# Annotate label defined by COCO
python3 annotate.py annotateCoco --image_directory=/path/to/the/image/directory/ --label=object_label_to_annotate --weights=/path/to/weights.h5 --displayMaskedImages=False
Note: --label=object_label_to_annotate should be in accordance to the COCO dataset labels. Refer COCO Dataset for more details.
Use pre-trained weights for the custom label. After finishing with the installation, we can run it directly from the command line as such:
# Annotate Custom
python3 annotate.py annotateCustom --image_directory=/path/to/the/image/directory/ --label=object_label_to_annotate --weights=/path/to/weights.h5 --displayMaskedImages=False
Note: --label=object_label_to_annotate should be a label for which the weights are provided.
Read the original post by Waleed Abdulla blog post about the balloon color splash sample where he explained the process starting from annotating images to training to using the results in a sample application.
The use train.py which is a modified version of balloon.py written by Waleed to support only the training part.
# Train a new model starting from pre-trained COCO weights
python3 customTrain.py train --dataset=/path/to/custom/dataset --weights=coco
# Resume training a model that you had trained earlier
python3 customTrain.py train --dataset=/path/to/custom/dataset --weights=last
