Note

This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.

Hugging Face

Dataset Card for 3dvs2026_papers#

image/png

This is a FiftyOne dataset with 176 samples.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/3dvs2026_papers")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details#

Dataset Description#

3DV 2026 Papers in FiftyOne is a multimodal dataset of all 177 accepted papers from the 13th International Conference on 3D Vision (3DV 2026), held in Vancouver, March 20–23, 2026. The dataset pairs per-page PNG images of each paper’s PDF (and supplementary material) with structured metadata and VLM-generated annotations, loaded into a FiftyOne grouped dataset for interactive exploration and analysis.

Each paper is represented as a FiftyOne group where every page image is its own slice. The dataset is designed to support conference navigation: browsing papers by topic, finding visually similar work, and surfacing the most distinctive or representative papers through embedding-based analysis.

  • Curated by: Harpreet Sahota

  • Language(s): English

  • License: The paper content belongs to the respective authors. Dataset structure, annotations, and code are released under MIT.

Dataset Sources#

Uses#

Direct Use#

  • Conference navigation: Browse and filter papers by research topic, session, affiliation, or contribution type in the FiftyOne App

  • Visual similarity search: Find papers with visually or semantically similar content using the Jina v4 embedding index

  • Paper discovery: Rank papers by uniqueness or representativeness to surface distinctive work or archetypal contributions in each area

  • Schedule planning: Cross-reference session metadata (oral_date, oral_start_time, poster_session) with topic and uniqueness scores to build a personalised conference agenda

Out-of-Scope Use#

  • This dataset is not intended for training machine learning models on paper content

  • The VLM-generated annotations (topic labels, affiliations, contribution flags) are model predictions and may contain errors — do not treat them as ground truth

Dataset Structure#

The dataset is a FiftyOne grouped dataset. Each group corresponds to one paper; each sample within the group is one page image.

Slice naming:

  • page_01, page_02, … page_NN — pages from the main PDF (8–22 pages per paper)

  • supp_page_01, supp_page_02, … — pages from supplementary material (1–83 pages; absent for papers without supplements)

Dataset statistics:

  • 176 papers with images (1 of 177 accepted papers had no available PDF)

  • 3,019 total page images across all slices

  • 22 oral papers, 154 poster-only papers

Fields on every sample:

Field

Type

Description

paper_id

string

Submission ID from OpenReview

title

string

Paper title

authors

list[string]

Author names

abstract

string

Full abstract text

affiliations

list[string]

Author institutions (VLM-extracted, ASCII-normalised)

poster_session

int

Poster session number (1–6)

poster_start_time

string

Poster session start time

oral_date

string

Date of oral presentation (null for poster-only papers)

oral_start_time

string

Oral session start time (null for poster-only papers)

pdf_link

string

OpenReview PDF URL

supplementary_link

string

OpenReview supplementary URL

project_page

string

Project page or GitHub URL (null if none found)

has_code

string

"Yes" / "No" derived from project_page

number_of_authors

int

Author count

topic_classifications

Classifications

VLM-assigned topic label (one of 10 categories)

introduces_dataset

string

"Yes" / "No" — paper introduces a new dataset

introduces_model

string

"Yes" / "No" — paper introduces a new model or architecture

introduces_method

string

"Yes" / "No" — paper introduces a new method or algorithm

introduces_benchmark

string

"Yes" / "No" — paper includes a benchmark or evaluation

jinav4_embeddings

vector

Per-page Jina v4 2048-dim embedding

all_page_embeddings_pooled

vector

Paper-level mean-pooled embedding (reference pages excluded)

jina_represent

float

Representativeness score (proximity to cluster centre)

jina_uniqueness

float

Uniqueness score (distance from nearest neighbours)

Topic taxonomy (10 categories):

  • 3D Reconstruction and Novel View Synthesis

  • 3D Generation and Editing

  • Human Avatars and Motion

  • Hand-Object and Human-Scene Interaction

  • Semantic 3D Understanding

  • Autonomous Driving & Outdoor Perception

  • Depth and Stereo Estimation

  • Dynamic and 4D Scenes

  • Geometric Vision Fundamentals

  • Novel Sensors and Specialized Domains

Dataset Creation#

Curation Rationale#

3DV 2026 received a record 404 submissions with a 43.8% acceptance rate, producing more good work than any attendee can absorb across 3.5 conference days. This dataset was created to support systematic, data-driven conference navigation: rather than relying on title skimming, attendees can explore the full research landscape through interactive visualisation, similarity search, and ranking.

Source Data#

Data Collection and Processing#

  1. Metadata scraping: Paper titles, authors, abstracts, PDF links, and supplementary links were collected from OpenReview for all 177 accepted papers. Session scheduling metadata (oral/poster dates and times) was extracted from the official 3DV 2026 conference programme.

  2. PDF rendering: Each accepted paper’s main PDF and supplementary material were downloaded from OpenReview and converted to per-page PNG images at 1700×2200px using pdf2image.

  3. FiftyOne grouped dataset construction: Images were loaded into a FiftyOne grouped dataset (parse_fiftyone_dataset.py) where each paper is a group and each page is a named slice. All metadata fields were attached to every sample.

Who are the source data producers?#

The paper content was produced by the authors of each accepted 3DV 2026 paper. Conference scheduling and programme information was produced by the 3DV 2026 organising committee (General Chairs: Manolis Savva, Angjoo Kanazawa, Christian Theobalt).

Annotations#

Annotation process#

All annotations were generated automatically using Qwen3.5-9B, a reasoning vision-language model, applied to the first page image of each paper (page_01) via the FiftyOne Model Zoo. Three annotation passes were run:

  1. Topic classification (classify operation): The model assigned each paper to exactly one of 10 predefined topic categories based on the visible title, abstract, and teaser figure. Prompt: structured JSON response [{"label": "topic_name"}].

  2. Affiliation extraction (vqa operation): The model extracted the list of author institutions appearing below the author names. ASCII normalisation was requested (e.g. ü→u) to avoid encoding inconsistencies. Response format: Python list of strings.

  3. Project page detection (vqa operation): The model identified any project page, GitHub repository, or supplementary site linked on the first page. Response format: URL delimited by triple backticks, or prose indicating no link was found.

  4. Contribution classification (vqa operation): The model determined whether each paper introduces a new dataset, model/architecture, method/algorithm, or benchmark. Response format: JSON object with boolean fields.

All model responses were parsed using two utility functions (parse_model_list_response, parse_model_url_response) that strip the <think>...</think> reasoning chain and extract the structured answer.

Embedding annotations were generated using Jina Embeddings v4 (jinaai/jina-embeddings-v4, retrieval task) via the FiftyOne Model Zoo, producing 2048-dim vectors for every page. Paper-level vectors were computed by mean-pooling per-paper page embeddings, excluding pages tagged reference-page (identified as a visual outlier cluster in the UMAP). Representativeness and uniqueness scores were computed using fob.compute_representativeness and fob.compute_uniqueness from the FiftyOne Brain.

Who are the annotators?#

Annotations were generated by Qwen3.5-9B (VLM, Qwen team / Alibaba Cloud) and Jina Embeddings v4 (Jina AI). No human annotators were used.

Personal and Sensitive Information#

The dataset contains publicly available author names, institutional affiliations, and paper abstracts sourced from OpenReview. This information was submitted voluntarily by paper authors as part of the academic peer review process and is publicly accessible on the OpenReview platform.

Bias, Risks, and Limitations#

  • VLM annotation accuracy: Topic labels, affiliation extractions, and contribution flags are model predictions. The topic taxonomy is coarse (10 categories) and papers near topic boundaries may be mislabelled. Affiliation strings may be incomplete or incorrectly normalised for some papers.

  • Reference page exclusion: Pages tagged reference-page were identified interactively by inspecting the UMAP visualisation. This process may have missed some reference pages or incorrectly excluded non-reference pages.

  • Geographic and institutional bias: The dataset reflects the accepted papers at one conference edition. Representation of institutions, countries, and research topics is determined by the 3DV 2026 submission and review process, not by this dataset.

  • PDF availability: 1 of 177 accepted papers had no PDF available on OpenReview at collection time and is absent from the image data.

Recommendations#

Topic labels should be treated as a navigational aid rather than a ground truth taxonomy. Users building on the VLM-generated annotations should validate a sample before relying on them for quantitative analysis.

Citation#

If you use this dataset or the accompanying code, please cite:

BibTeX:

@misc{sahota2026_3dvs_fiftyone,
  author       = {Sahota, Harpreet},
  title        = {3DV 2026 Papers in FiftyOne},
  year         = {2026},
  howpublished = {\url{github.com/harpreetsahota204/awesome_3dvision_2026_conference}},
  note         = {Dataset of accepted papers from the 13th International Conference on 3D Vision}
}

Dataset Card Authors#

Harpreet Sahota