TensorFlow Object Detection

Introduction

Annotation

To draw bounding boxes using a GUI, I recommend sloth.

To install the GUI, you need

• Python 3.5
• PyQt4

Use Conda to get those dependencies work togeter

conda create -n slot_env python=3.5
source activate slot_env
pip install git+https://github.com/cvhciKIT/sloth.git
conda install -c anaconda pyqt=4.11.4

If sloth is installed, you can type the executable sloth to lanch the GUI.

Configure Sloth for custom labels

We need to configure Sloth so that we can create our label set.

We can create a configuration file as a Python script. A configuration is a list of dictionaries.

For example, we want to draw rectangle bounding boxes for two classes cat and dog

LABELS = (
    {"attributes": {"type":  "rect",
                    "class": "cat",
                    },
     "item":     "sloth.items.RectItem",
     "inserter": "sloth.items.RectItemInserter",
     "text":     "cat"
    },
    {"attributes": {"type":  "rect",
                    "class": "dog",
                    },
     "item":     "sloth.items.RectItem",
     "inserter": "sloth.items.RectItemInserter",
     "text":     "dog"
    },
)

Save the config as my_config.py, then launch Sloth GUI with custom labels

sloth --config my_config.py

The annotation will be saved as json. Note that each json file consists of one or more images.

For example, if you annotate two images, the first images is annotated with a dog and a cat, and the second with just a cat, the output json can be like

[
    {
        "class": "image",
        "filename": "image1.jpg",
        "annotations": [
            {
                "class": "cat",
                "height": 60.0,
                "width": 46.0,
                "y": 105.0,
                "x": 346.0
            },
            {
                "class": "dog",
                "height": 58.0,
                "width": 56.0,
                "y": 119.0,
                "x": 636.0
            }
        ]
    },
    {
        "class": "image",
        "filename": "image2.jpg",
        "annotations": [
            {
                "class": "cat",
                "height": 60.0,
                "width": 46.0,
                "y": 105.0,
                "x": 346.0
            },
        ]
    },
]

More to come