3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions (1603.08182v3)

Abstract: Matching local geometric features on real-world depth images is a challenging task due to the noisy, low-resolution, and incomplete nature of 3D scan data. These difficulties limit the performance of current state-of-art methods, which are typically based on histograms over geometric properties. In this paper, we present 3DMatch, a data-driven model that learns a local volumetric patch descriptor for establishing correspondences between partial 3D data. To amass training data for our model, we propose a self-supervised feature learning method that leverages the millions of correspondence labels found in existing RGB-D reconstructions. Experiments show that our descriptor is not only able to match local geometry in new scenes for reconstruction, but also generalize to different tasks and spatial scales (e.g. instance-level object model alignment for the Amazon Picking Challenge, and mesh surface correspondence). Results show that 3DMatch consistently outperforms other state-of-the-art approaches by a significant margin. Code, data, benchmarks, and pre-trained models are available online at http://3dmatch.cs.princeton.edu

• The paper introduces 3DMatch, a novel deep learning framework for extracting local geometric descriptors from RGB-D scans.
• It demonstrates improved matching accuracy over traditional hand-engineered features by leveraging volumetric convolutional networks.
• The method enhances applications in surface reconstruction and robotic scene understanding by providing reliable feature correspondences.

Author Guidelines for CVPR Proceedings

This paper provides a comprehensive guide for authors preparing manuscripts for the Conference on Computer Vision and Pattern Recognition (CVPR). It outlines the essential steps and formatting instructions required for submissions, ensuring uniformity and facilitating a seamless review process.

Abstract Guidelines

The abstract is required to be in italicized text, fully justified, and placed at the top of the left-hand column below the author and affiliation information. The guidelines specify that the abstract should be concise, using a 12-point Times boldface title and 10-point single-spaced type for the text. Observance of these formatting rules is crucial for authors to conform to the CVPR standards.

Main Body Instructions

Language and Paper Length

The paper mandates that all contributions should be in English. There is a strict limit on paper length: excluding references, the paper must not exceed eight pages. Authors are warned that any deviation, such as altered margins or formatting to extend the length, will result in the paper not being reviewed. This rigorous enforcement ensures fairness and maintain the standard format for all submissions.

Blind Review

Anonymization for blind review is emphasized, clarifying common misconceptions. Authors are advised not to use words like "my" or "our" when citing their work. Proper anonymization practices include referring to previous work without revealing author identity directly. This practice aids in maintaining a fair and unbiased review process.

Mathematics and Figures

The document stresses the importance of numbering all sections and displayed equations for ease of reference. This helps future readers locate specific content accurately. Furthermore, figures and tables must be appropriately sized and centered, with captions in 9-point Roman type to maintain visual coherence. Authors are encouraged to use LaTeX commands like \includegraphics for consistent figure inclusion.

Formatting and Submission

The guide specifies rigorous formatting details for the text, including two-column layout, Times font usage, and precise margin dimensions. Key instructions include:

• Main text in 10-point Times, single-spaced.
• First-order headings in 12-point boldface.
• Second-order headings in 11-point boldface.
• Footnotes and references in 8 and 9-point Times, respectively.

The meticulous formatting instructions ensure that all manuscripts appear uniform, which streamlines both the review process and eventual publication.

Additional Instructions

Authors are provided with examples demonstrating correct and incorrect use of self-citations under blind review. The examples elucidate how to reference one’s previous work without compromising anonymity. There is also advice on handling dual submissions and the integration of technical reports as supplementary material.

Practical Implications

For practical submissions, adhering to these guidelines is essential. The strict page limit and formatting rules influence the way authors present their research, encouraging conciseness and clarity. The blind review process helps maintain objectivity and fairness, which are fundamental to the integrity of the conference.

Future Implications

Looking ahead, the stringent formatting and review guidelines set by CVPR could encourage other conferences to adopt similar standards to improve submission quality and review efficiency. This consistency can lead to broader adoption of good practices in research dissemination in the computer vision community and beyond.

Conclusion

This guide offers essential directives for the preparation of CVPR submissions, covering aspects from language and length to figures and references. By following these instructions, authors can ensure their manuscripts meet the required standards, facilitating a fair and efficient review process and contributing to the high quality of the CVPR proceedings. The impact of these guidelines extends beyond the conference, promoting best practices in academic publishing.