This code produces object candidates (in bounding boxes) from RGBD frames.

If you use this software, please cite the following paper:

Asako Kanezaki and Tatsuya Harada. 3D Selective Search for Obtaining Object Candidates. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2015.
(pdf)

ROS indigo/jade
http://wiki.ros.org/

For example,
$ roslaunch freenect_launch freenect.launch
In this case, you will get point cloud in rostopic /camera/depth_registered/points and color images in rostopic /camera/rgb/image_color.

$ rosrun selective_search_3d selective_search_3d -v 0.02 -d 1.0 points:=/camera/depth_registered/points
In this case, you will get bounding boxes of object candidates within depth range = 1.0m. The voxel resolution is set to 0.02m.

$ rosrun selective_search_3d show_bbox -n 5 image:=/camera/rgb/image_color
or
$ python show_bbox.py -n 5 image_topic:=/camera/rgb/image_color
In this case, five bounding boxes at maximum are shown.

$ rosrun selective_search_3d print_bbox3D -n 5
or
$ python print_bbox3D.py -n 5
In this case, five bounding boxes at maximum are printed.

Please install Fast R-CNN according to https://github.com/rbgirshick/fast-rcnn

Then, specify the path of the directory that you cloned Fast R-CNN into fast_rcnn_path in do_fast-rcnn.py.
$ python do_fast-rcnn.py --net caffenet --conf 0.7 image_topic:=/camera/rgb/image_color
Please change the threshold of confidence (0.7) on demand.

$ rosrun selective_search_3d selective_search_3d_offline -v 0.02 -p cloud.pcd -o bbox.txt
In this case, you will get bounding boxes of object candidates written into bbox.txt. The voxel resolution is set to 0.02m.

BSD