This python based toolbox provides demos on how to use FUTURE3D dataset.

• Blender 2.7.9b (https://www.blender.org/download/releases/2-79/)
• Open3d
• Scipy = 1.4.1
• Numpy
• OpenCV

Default blender python environment is different with your custom python environment, so you may need to install required python modules(e.g. Numpy) for blender python following totorials below:

• http://www.codeplastic.com/2019/03/12/how-to-install-python-modules-in-blender/

• https://blender.stackexchange.com/questions/5287/using-3rd-party-python-modules

If above tutorials do not work for you, you can add custom environment path into system path in python script as below : (more details please refert to demo_render_object_mv_with_corrs_model.py)

# add custom python env
sys.path += ['/home/shunming/software/envs/dnn/lib/python3.6/site-packages'] # Please change into your path

With provided models and pose, you can generate idmap using blender (Instance Segmentation only).

bash scripts/demo_render_scene_image.sh

bash scripts/demo_project_scene.sh

Please normalize models into (-1, 1) with norm_model.py, if using default settings.

This script support rendering for textured image, normal image, depth image and silhouette image. The type of depth image is RGBA, so you need to convert it into RGB to get correct relative depth values (range from 0 to 255).

Note: If you want to get true depth value, please use 'OPEN_EXR' instead of 'PNG' in blender.

bash demo_render_object_mv_with_texture.sh

If provided multi-view images are not enough for your training(default: 12 views), you can run this script to generate more data with random viewpoints. Same as texture rendering, this script also supports depth image, normal image, and silhouette image

bash demo_render_object_with_texture_random.sh

Generate four different types of rendering image as well as relevant obj model with aligned pose. This would help to create your own dataset, especially for the reconstruction task.

bash scripts/demo_render_object_mv_without_texture.sh

With 3D models, 2D rendering images, and relevant pose info, you can project 3D points into 2D image plane using this script - valid for both simple multi-view images(12 views) and random view images.

bash scripts/demo_project_mv.sh

Giving a depth map(type-exr, true value), reproject 2d points into camera coordinate.

bash scripts/demo_reprojection_mv.sh

bash scripts/demo_render_object_with_texture_pose.sh

We borrow some codes from 3D-R2N2 to develop our toolbox.

If you use this toolbox for your research, please agree the terms and conditions in the "3D-FUTURE Data Sets Use License Agreement": English Version, 中文版, and cite our paper(will be publised soon).

If you have any questions, please contact [email protected]