# CITATIONS.md

External datasets, models, and methodologies this package depends on.
Bibliographic entries below are abbreviated; the full bibtex ships with
the camera-ready paper.

## Datasets

### Objaverse and Objaverse-LVIS

Deitke et al., "Objaverse: A Universe of Annotated 3D Objects." CVPR 2023.
We use the Objaverse-LVIS subset (46,205 evaluation UIDs, 1,156 canonical
classes). UIDs are publicly redistributable; the evaluation split (UIDs +
categories + standard OpenShape data-layout paths) ships at
`data/splits/lvis_eval.json`. The 1,156 canonical class names are
extracted on-the-fly from that file by `code/utils/load_lvis_classes.py`
(no separate names file is bundled).

### Cap3D

Luo et al., "Scalable 3D Captioning with Pretrained Models." NeurIPS 2023.
We use the Cap3D-full caption set (~874K captions over the Objaverse
training pool) for the own-caption stratification analysis and as the
contaminant whose presence the counterfactual prune masks target. Captions
themselves are not redistributed in this package; reviewers download them
from the Cap3D HuggingFace dataset. We do ship the per-UID per-filter-family
boolean flags derived from the captions.

### Sketchfab captions (sf_name_desc, sf_tags)

We additionally use Sketchfab-name-and-description and Sketchfab-tags
captioner variants from the original Objaverse metadata for the
captioner-corpus sweep (Table 1, captioner row). These are part of the
Objaverse metadata release; see Objaverse paper.

### LVIS class taxonomy

Gupta et al., "LVIS: A Dataset for Large Vocabulary Instance Segmentation."
CVPR 2019. We use only the canonical class-name string list (1,156 names);
these are derived from the `category` field of `data/splits/lvis_eval.json`.

### ScanObjectNN (OBJ_ONLY and PB_T50_RS variants)

Uy et al., "Revisiting Point Cloud Classification: A New Benchmark Dataset
and Classification Model on Real-World Data." ICCV 2019. Used for
cross-benchmark NN-proxy analysis (Table 2 cross-benchmark extension).
Test splits are part of the ScanObjectNN release. The per-object
cross-benchmark prediction parquet is **not bundled** with this package
(see CHANGELOG / REPRODUCE caveat: reviewer-produced via
`code/uni3d_replication/`); reviewers download original point clouds and
regenerate predictions from the shipped scripts.

### ModelNet40

Wu et al., "3D ShapeNets: A Deep Representation for Volumetric Shapes."
CVPR 2015. Test split used for the cross-benchmark NN-proxy analysis
and for the MN40 insulation column of Table 3. Same arrangement as
ScanObjectNN.

## Models

### OpenShape

Liu et al., "OpenShape: Scaling Up 3D Shape Representation Towards
Open-World Understanding." NeurIPS 2023. The PointBERT ViT-G/14 RGB
backbone is the primary subject of our counterfactual retraining
analysis. Pretrained weights are downloaded from the OpenShape release;
SHA-256 of the weights is recorded in `data/audit/G0_FINAL.md`.

### PointBERT

Yu et al., "Point-BERT: Pre-training 3D Point Cloud Transformers with
Masked Point Modeling." CVPR 2022. The point-cloud encoder backbone used
by OpenShape.

### Uni3D

Zhou et al., "Uni3D: Exploring Unified 3D Representation at Scale." ICLR
2024. Used for the +/-LVIS gap replication and per-stratum decomposition
(Section 4.4.3). Pretrained weights (both no-LVIS and with-LVIS variants)
are downloaded from the Uni3D release.

### OpenCLIP ViT-bigG-14

Ilharco et al., "OpenCLIP." Used as the text encoder for zero-shot
classification (laion2b_s39b_b160k variant). Downloaded automatically by
`open_clip_torch`.

### CLIP (ViT-G/14)

Radford et al., "Learning Transferable Visual Models From Natural Language
Supervision." ICML 2021. Original CLIP for context; we use the OpenCLIP
ViT-bigG-14 variant in practice.

## Methodology

### Mantel-Haenszel pooled risk difference (Greenland estimator)

Greenland and Robins, "Estimation of a common effect parameter from
sparse follow-up data." Biometrics 1985. Used throughout for within-class
pooled estimation.

### Sato closed-form CI

Sato, "Confidence intervals for effect parameters common in cancer
epidemiology." Environmental Health Perspectives 1990. Used as the
secondary CI estimator per VERDICT Required Revision #3 (cluster
bootstrap is primary).

### Cluster bootstrap

Efron and Tibshirani, "An Introduction to the Bootstrap." Chapman &
Hall, 1993. Used at the class level (B=2000) per VERDICT Required
Revision #3.

### Cohen's h effect size

Cohen, "Statistical Power Analysis for the Behavioral Sciences." Erlbaum,
1988.

## Software dependencies (programmatic)

Pinned versions in `pyproject.toml` / `uv.lock`. Headline:

- `torch==2.4.1` (Paszke et al., NeurIPS 2019)
- `transformers` >= 4.40 (Wolf et al., EMNLP 2020 demo)
- `open_clip_torch` >= 2.24
- `scipy` >= 1.11
- `pandas` >= 2.1, `pyarrow` >= 14.0
- `nltk` >= 3.8 (for word-boundary tokenization)

## This work

The BMVC 2026 submission "Two Axes of Caption Contamination in
Open-Vocabulary 3D Classification." Author names and institutional
affiliations omitted for double-blind review per BMVC policy. The
camera-ready bibliography will replace this paragraph.

---

## Notes on data licensing

- Objaverse: ODC-By 1.0.
- Cap3D: CC BY 4.0 (per the Cap3D paper).
- ModelNet40: research-use license; see original paper.
- ScanObjectNN: see project page.
- OpenShape / Uni3D weights: research-use license; see respective releases.

This supplementary package is released under MIT (see `LICENSE`). Use of
the external datasets and models above is subject to their original
licenses, which take precedence for any redistributed artifacts derived
from them.