PPAT¶

PPAT (Point Patch Transformer) maps a colored point cloud into an OpenShape CLIP embedding. Shapes can then be ranked against text labels without training a task-specific classifier. WarpConvNet loads the released OpenShape PointBERT checkpoints directly.

PPAT is separate from Uni3D. Each family requires text features from its own paired text encoder; raw cosine scores are not comparable across the two embedding spaces.

Supported models¶

Variant	Up axis	Output width	Text space
openshape-pointbert-vitb32-rgb	Y	512	CLIP ViT-B/32
openshape-pointbert-vitl14-rgb	Y	768	CLIP ViT-L/14
openshape-pointbert-vitg14-rgb	Z	1,280	CLIP ViT-bigG/14

All variants accept 10,000-point XYZ+RGB inputs and load their matching OpenShape weights. B32 and L14 are Y-up; bigG is Z-up. Rotate the shape into the checkpoint's up-axis convention before normalization. The model and normalize_point_cloud do not rotate coordinates.

Usage¶

Install the shared demo dependencies for Hugging Face downloads and mesh loading:

pip install -e ".[demo]"

import torch

from warpconvnet.models.ppat import load_openshape_pointbert, normalize_point_cloud

model = load_openshape_pointbert(
    "openshape-pointbert-vitg14-rgb",
    device="cuda",
)

# xyz and rgb have shape (batch, points, 3). xyz is already z-up.
xyz = normalize_point_cloud(xyz)
points = torch.cat((xyz, rgb), dim=-1).permute(0, 2, 1).contiguous()

with torch.inference_mode():
    embedding = model(points.cuda())

The model input has shape (batch, 6, points), with XYZ in channels 0-2 and RGB in channels 3-5. RGB values are in [0, 1]. normalize_point_cloud centers XYZ and scales it to the unit ball; it does not rotate the input.

load_openshape_pointbert downloads and strictly loads published weights. Set checkpoint_path for local weights or use build_openshape_pointbert to create an uninitialized model.

The bundled mesh demo downloads the bigG model and prompt-averaged LVIS text features:

python examples/ppat_demo.py

It prints labels ranked by cosine similarity. These values are relative scores, not probabilities.

Neighborhood backends¶

Backend	Device	Behavior
cumsum	CPU/CUDA	Reference-compatible point order and default
dense	CPU/CUDA	Dense reference implementation; bit-identical to cumsum
warp	CUDA	Exact cell-list radius search; lower memory but not faster
ball_cuda	CUDA	Exact spherical rule with direct-distance boundary rounding
voxel:SIZE	CPU/CUDA	Approximate 3x3x3 voxel neighborhood
voxel_cuda:SIZE	CUDA	Fused approximate voxel neighborhood
l1, linf, random, knn	CPU/CUDA	Optional approximate selectors

warp and ball_cuda follow the same radius rule as the reference path but compute distance directly. Floating-point rounding at the radius boundary can therefore differ from cumsum. Approximate backends can change model accuracy and should use the same normalization and configuration during evaluation and deployment.

Accuracy and performance¶

The OpenShape checkpoint table reports 46.8% Objaverse-LVIS zero-shot top-1 accuracy for pointbert-vitg14-rgb.

The latency measurements below use the supported WarpConvNet path in BF16 on an NVIDIA RTX PRO 6000 Blackwell (sm_120).

Batch 32, 10,000 points	Latency
Whole model	17.06 ms
Transformer	4.69 ms

Ball query, B=96, N=10,000, S=384, K=64	Time	Peak memory
voxel_cuda:0.1	0.53 ms	0.10 GiB
ball_cuda	0.79 ms	0.10 GiB
cumsum	20.25 ms	3.09 GiB
dense	25.30 ms	9.63 GiB

voxel_cuda:0.1 is approximate. ball_cuda, cumsum, and dense implement the same spherical neighborhood rule, subject to the floating-point boundary behavior described above.

See the package README for the full input contract and demo options.