MSCOCO

COCO[2] refers to a common object in context and is a dataset for object recognition, with 80 categories and 330K images. After Pascal VOC'12, this became a popular benchmark for training and evaluating the system. The dataset can be downloaded from http://cocodataset.org/#download. 

In order to read the data and use it for applications, there is an API available at https://github.com/cocodataset/cocoapi which needs to be downloaded.  To get started, we can use the API provided, as follows:

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
make

This will install the Python API to read the coco dataset. 

Many models available online for object detection or image segmentation are first trained on this dataset. If we have specific data that has different object categories than in the MSCOCO dataset, a more common approach that we will see in Chapter 5, Convolution Neural Networks and in Chapter 6, Feature- Based Object Detection, is to first train a model on an MSCOCO dataset and use a part of the trained model and re-train on a new dataset.