This is part of my Winter Project at NWU.

Use a depth camera to detect objects on table and get their mask in 2D image.

1. Get the point cloud from rgbd image (rgbd image = color + depth image)
2. Detect and remove plane (table surface). The rest clouds are the objects.
3. Do clustering. Each cluster is an object.
4. Project 3D points onto 2D image to obtain the mask of objects.
5. Do a dilation and erosion to fill holes, and get the final mask.

There are 3 choices for the input RGBD-D images:
(1). Realsense camera.
(2). AsusXtion camera.
(3). Load from file.

Search for "ObjectDetectorFromRGBD". There are several parameters such as "zmin", "zmax", etc., which are for range filtering of the input cloud, and "voxel_size" for downsampling the input cloud.

Search for "node_detect_object_from_cloud". There are several paramters related to removing plane and do clustering. You can modify them.

Based on the data source you choosed in step 1, run one of the follows:

$ roslaunch mask_objects_from_rgbd open_realsense.launch

or:

$ roslaunch mask_objects_from_rgbd open_xtion.launch

or:

Do nothing if you choose to load from file. But remember to modify the paths of data source in src/n0_fake_rgbd_image_publisher.py.

$ roslaunch mask_objects_from_rgbd main.launch

The result is shown in rviz, similar to the demo image above.

You can also press 's' to save the results. They will be saved into data/ folder as follows:

data
└── 03-21-14-03-22-686
    ├── clouds
    ├── depth
    ├── image
    ├── mask
    ├── objects
    └── resimg

(Notes: I'm using cv2.waitKey(), so you need to click the showed image, and then press 's'.)