ShapenetRender_more_variation

A new shapenet rendering 2D image dataset that also contains deph map, normal map and albedo map.

Please cite our paperDISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction (NeurIPS 2019) if you plan to download the rendered images or use our code to render by yourself.

@inProceedings{xu2019disn,
  title={DISN: Deep Implicit Surface Network for High-quality Single-view 3D Reconstruction},
  author={Xu, Qiangeng and Wang, Weiyue and Ceylan, Duygu and Mech, Radomir and Neumann, Ulrich},
  booktitle={NeurIPS},
  year={2019}
}

Code contact: Qiangeng Xu* and Weiyue Wang*

Also please cite Shapenet's original paper as well.

Dataset Intro:

The categories included are:

        "watercraft": "04530566",
        "rifle": "04090263",
        "display": "03211117",
        "lamp": "03636649",
        "speaker": "03691459",
        "cabinet": "02933112",
        "chair": "03001627",
        "bench": "02828884",
        "car": "02958343",
        "airplane": "02691156",
        "sofa": "04256520",
        "table": "04379243",
        "phone": "04401088"
        }

Our rendering is based on the convention of 3DR2N2's 2d image rendering.

• Our script support both shapenet v1 and v2, for v2 you need to change render_blender.py's random range since some model always get out of the field of view. The tar files we provide is rendered on v1.

• Each model object has 36 easy views and 36 hard views.(3DR2N2 has 24 easy views)

• Each view of each model object, we have albedo, depth, normal and RGB images.(3DR2N2 has only RGB images)

• different from 3DR2N2, our resolution is 224 224 instead of 137 137

• All the object is absolutely inside the field of view.

• Camera is looking at origin, but in hard folder, we randomly shift the center of the model at (x_rand,y_rand,z_rand)

• We set camera Pitch to 0 degree since in most case in real world, the floor is flat.(We assume camera to the origin is along Z axis, the rotation along Z is 0 degree)

albedo	RGB	Depth	normal

In each folder, there is a meta file: rendering_metadata.txt: each line represent a parameter:

Dataset download:

image.tar

albedo.tar

depth.tar

normal.tar

Or you can run the generation script by yourself :

  install blender 2.79 and go to its python3.5m to install pip3, then install numpy and opencv

  python -u render_batch --model_root_dir {model root dir} --render_root_dir {where you store images} --filelist_dir {which models you want to render} --blender_location {you} --num_thread {10} --shapenetversion {support v1, v2} --debug {False}

Transformation matrix calculation:

Please refer to cam_read.py