Skip to content

README.md

Latest commit

 

History

History
74 lines (55 loc) · 2.73 KB

README.md

File metadata and controls

74 lines (55 loc) · 2.73 KB
Raw

ShapeNetPU Data

Download ShapeNetPU from Hugging Face to data/ShapeNetPU. Commands are in Data preparation.

Raw Samples

ShapeNetPU/
|-- train/<category>/<sample>/
|   |-- input/sparse_xyz.ply
|   |-- input/rendering/               # PNG views and renderings.txt
|   `-- gt/gt_xyz.ply
`-- test/<category>/<sample>/
    |-- input/sparse_xyz.ply
    |-- gt/gt_xyz.ply
    `-- gt/gt_mesh.obj

Each pair contains 2,048 sparse points and 8,192 GT points. Coordinates support binary_little_endian PLY and NPY. The basenames sparse_point_cloud and upsampled_point_cloud are also recognized.

Prepared Features

Path within a sample Shape / dtype Usage
input/sparse_local_feat.npy [2048,128], float32 Training and inference
input/vggt_multiview_features.npz patch_tokens=[8,32,512], float16 Training
input/vggt_feature_meta.json View order, export settings and hashes Feature metadata
gt/gt_feat.npy [8192,128], float32 Training

Feature rows follow point order. The VGGT NPZ key is patch_tokens; its memmap array name is vggt_image_cond. Export commands are in Feature preparation.

Images are read from input/rendering/, input/images/, or a ShapeNetRendering/<category>/<sample>/rendering/ tree inside or alongside the dataset. View order follows renderings.txt when present, otherwise sorted filenames. Use --render_root for an external rendering tree.

Data Loading

Mode Data source
Source files Per-sample point clouds and prepared NPY/NPZ features
Memmap (recommended) Prepacked, memory-mapped NPY arrays

Memmap reduces small-file I/O and point-cloud parsing. The loader prefers .controlflow3d_memmap/<split>/manifest.json when present and reads source files otherwise. Set CONTROLFLOW3D_DISABLE_MEMMAP_CACHE=1 to force source-file loading.

Build the optional cache after feature preparation:

python utils/build_memmap_dataset_cache.py \
  --dataset_root data/ShapeNetPU --split all \
  --index_mode part --feature_dtype float16

Training caches include coordinates, sparse/dense features and image conditions; test caches include coordinates and sparse features. Add --include_image_cond for image-conditioned testing. Rebuild with --overwrite after changing features.

  • part: contiguous point blocks, without requiring PLY patch indices.
  • patch: uses PLY patch indices; select --index_mode patch for both packing and training.

Additional options: python utils/build_memmap_dataset_cache.py --help.