point sampling code for IM-NET

The code samples 3 levels of resolutions for coarse-to-fine training: 16³, 32³, 64³. But the actual sampled points are from 256³ voxels.

For 16³ resolution, we sample 16×16×16 = 4096 points. We first divide the ground truth 256³ voxels into 16³ cubes each containing 16³ voxels. For each cube, if it contains at least one "inside" voxel, we randomly sample one of the "inside" voxels and use the center of that voxel as the sample point. Otherwise we randomly sample one "outside" voxel.

For 32³ and 64³ resolution, we sample 4096 and 16384 points respectively. We sample those voxels adjacent to the surface of the shape. Please refer to the code for details.

Usage

Download the original voxel models from HSP. Download the rendered views from 3D-R2N2.

step 1: change the data directories in 1_check_hsp_mat.py and run it. Because a few voxel files are corrupted, you need to manually remove them. Their names will be printed out by the script. Also, remove 4a32519f44dc84aabafe26e2eb69ebf4 from category rifle (04090263).

step 2: change the category name and the data directories in 2_gather_256vox_16_32_64.py and run it for each category. Use 2_test_hdf5.py to check if the output hdf5 is correct.

step 3: change the category name and the data directories in 3_gather_img.py and run it for each category.

step 4: run 4_gather_all_vox_img_test.py, 4_gather_all_vox_img_train.py to get a hdf5 file containing all categories, or run 4_gather_vox_img_test.py, 4_gather_vox_img_train.py to get hdf5 files containing individual categories.

If you have prepared your own voxel data, please modify the code accordingly. The only difference is the code segment for reading voxels.