eylulpelinkilic/Colored_Point_Clouds

Colored point-cloud completion — PoinTr-protocol occlusions

Colored ground truth and occluded partial inputs for colored point-cloud completion, built from textured ShapeNetCore meshes. Three categories, 599 models, 1797 partials.

What makes this different from the usual completion sets: the ground truth carries per-point color sampled from the mesh texture, and it is de-speckled before sampling, so the color is actually correct rather than plausible-looking.

Contents

gt_s3/<synset>/<model_id>.ply                          colored GT, 8192 pts, xyzrgb
gt_s3/gt_manifest.csv
occluded_occ/<difficulty>/<synset>/<model_id>.ply          partial input
occluded_occ/<difficulty>/<synset>/<model_id>_missing.ply  removed region
occluded_occ/manifest.csv

difficulty ∈ simple (25 % removed) · moderate (50 %) · hard (75 %).

For every model partial ∪ missing = gt exactly — the two files are a partition of the same 8192 points, so the removed region is itself labelled ground truth, not an approximation.

synset	category	models	partials
02691156	airplane	200	600
02958343	car	199	597
03001627	chair	200	600
total		599	1797

One car model (15fcfe91d44c0e15e5c9256f048d92d2) is absent — its texture is missing from the source ShapeNetCore archive.

How the ground truth was built

A naive colored mesh sample is wrong, not merely noisy: ShapeNet meshes carry double-sided faces, so a sampler picks back-face colors and produces speckle. The fix is part of the pipeline:

1. pymeshlab ambient occlusion → per-face visibility.
2. Dual-face removal — among faces sharing the same three vertices, keep the most visible one (Lazzarotto & Ebrahimi, EPFL/MMSPG, arXiv:2201.06935). On a single test model this alone moves color ΔE from 13.1 → 0.08.
3. Dense color sample of the cleaned mesh (CloudCompare, 2¹⁷ points).
4. Color-aware farthest-point sampling → fixed 8192 points.

De-speckling has to run on the mesh before sampling, so the order is not optional.

Occlusion protocol

PoinTr / ShapeNet-55 style, ported from seprate_point_cloud: draw a random unit viewpoint v, sort points by distance to v, remove the nearest crop_ratio · N.

This is not the PCN protocol. PCN back-projects depth renders from 8 viewpoints and models a real range scan; this one removes a known fraction of the surface and is meant for controlled difficulty sweeps. They are different task definitions — a benchmark wants both, and results are not interchangeable.

occluded_occ/manifest.csv records synset, model, difficulty, crop_ratio, view, seed, n_partial, n_missing, gt_ply. The seed makes every partial reproducible from the GT alone.

Loading

import open3d as o3d, numpy as np

def load(p):
    pc = o3d.io.read_point_cloud(p)
    return np.asarray(pc.points), np.asarray(pc.colors)   # (N,3) xyz, (N,3) rgb in [0,1]

xyz_in,  rgb_in  = load("occluded_occ/moderate/02691156/<model>.ply")          # input
xyz_tgt, rgb_tgt = load("occluded_occ/moderate/02691156/<model>_missing.ply")  # target
xyz_gt,  rgb_gt  = load("gt_s3/02691156/<model>.ply")                          # input ∪ target

Suggested metric

Color error on the missing region only — mean ΔE in CIELAB between prediction and ground truth. Evaluating over the whole cloud hides the task: the visible part is given.

Note that ΔE is a distortion metric and provably rewards the conditional mean, so it favours averaging baselines over generative ones (distortion–perception trade-off, Blau & Michaeli, 2018). Report a distribution metric alongside it.

Provenance and license

Derived from ShapeNetCore v2 meshes and ShapeNet-Part model lists. Only derived point clouds are published — the source meshes are not redistributed. ShapeNet's original terms govern anything derived from it; read them before redistributing.

Produced by scripts/build_gt.py → scripts/build_occluded.py.

Related

A companion dataset by the same project covers the PCN protocol (8 depth-render views per model) over the same three categories: efeyenice/pc-completion-data.