Deep Learning with Your Own Image Dataset

This page shows the overview of how to start deep learning with your own image dataset.

1. Collect raw images

Collect available images from Web or take pictures by yourself.

2. Annotate images

Label each image or each region in the images.

• Annotate images with labelme

3. Train neural network

Choose one network model according to your task.

There are some training scripts in jsk_perception package.

Training for semantic segmentation

• Train FCN (Fully Convolutional Network)

Training for instance segmentation

• Train Mask-RCNN

Training for object detection

• Train SSD

4. Infer with trained network

Now you can use your trained network to infer things as a ROS node.

Nodes for semantic segmentation

• fcn_object_segmentation.py

Nodes for instance segmentation

• mask_rcnn_instance_segmentation.py

Nodes for object detection

• ssd_object_detector