[Paper Review] A Large Dataset of Object Scans

Sungjoon Choi, Qian-Yi Zhou+2 more|arXiv (Cornell University)|Feb 8, 2016

3D Surveying and Cultural Heritage6 references116 citations

TL;DR

The paper presents a large, public-domain dataset of over 10,000 3D object scans collected by 70 non-expert operators using consumer-grade mobile RGB-D setups, plus reconstructed models and analysis of scan diversity and reconstruction success.

ABSTRACT

We have created a dataset of more than ten thousand 3D scans of real objects. To create the dataset, we recruited 70 operators, equipped them with consumer-grade mobile 3D scanning setups, and paid them to scan objects in their environments. The operators scanned objects of their choosing, outside the laboratory and without direct supervision by computer vision professionals. The result is a large and diverse collection of object scans: from shoes, mugs, and toys to grand pianos, construction vehicles, and large outdoor sculptures. We worked with an attorney to ensure that data acquisition did not violate privacy constraints. The acquired data was irrevocably placed in the public domain and is available freely at http://redwood-data.org/3dscan .

Motivation & Objective

• Provide a large-scale, diverse 3D object scan dataset collected in uncontrolled, real-world environments to reflect consumer deployment conditions.
• Assess the practicality and challenges of object reconstruction pipelines on non-expert, mobile-scanning data.
• Release both raw scans and reconstructed models to support research in 3D reconstruction, modeling, and related tasks.
• Document data collection procedures, object diversity, and category-wise reconstruction outcomes to guide future dataset design.

Proposed method

• Assembled ten lightweight mobile scanning setups combining netbooks, RGB-D cameras, and carrying cases for handheld data capture.
• Used a custom scanning application with live color feed and distance-based color coding to guide operators and ensure coverage of object surfaces.
• Recruited 70 operators via campus outreach; provided tutorials and guidelines; compensated by recording time to encourage thorough scanning.
• Collected videos of objects chosen by operators, under IRB approval and with privacy considerations, and pruned inadequate scans before release.
• Reconstructed 3D models using a hybrid odometry pipeline that combines ICP-based geometric alignment with RGB-D photometric alignment: E(Ti)=EICP(Ti)+λERGBD(Ti).
• Provided a dataset including raw RGB-D scans and a subset of reconstructed models; noted limitations of the reconstruction pipeline for large objects and handheld scans.

Experimental results

Research questions

• RQ1 What is the scale and diversity of object scans obtainable when non-experts use consumer-grade mobile scanning setups in real-world environments?
• RQ2 How do reconstruction pipelines perform on consumer-grade, in-the-wild scans, and what are their failure modes?
• RQ3 What is the category-wise success rate and overall feasibility of reconstructing 3D models from such scans?
• RQ4 How does scan modality (handheld vs stationary) influence data quality and reconstruction outcomes?

Key findings

• The dataset contains over 10,000 dedicated object scans with an average scan length of 77 seconds, yielding more than 23 million RGB-D images.
• Objects span a wide range of categories; vehicles account for ~13% of scans and chairs/tables ~10% combined across 44 categories with at least 44 scans each (h-index = 44).
• 1,781 sequences were processed by the reconstruction pipeline; 969 were lost due to drift or fast motion, 812 were kept after filtering, and 398 models passed qualitative inspection for public release.
• The reconstruction pipeline achieved an overall success rate of 22% across nine object categories (Chair, Table, Trash container, Bench, Plant, Sign, Bicycle, Motorcycle, Sofa).
• Three large objects were reconstructed with a high-fidelity pipeline that includes loop closure and global optimization, beyond the standard pipeline applied to the bulk of data.