Accurate Instance-Level CAD Model Retrieval in a Large-Scale Database

Abstract: We present a new solution to the fine-grained retrieval of clean CAD models from a large-scale database in order to recover detailed object shape geometries for RGBD scans. Unlike previous work simply indexing into a moderately small database using an object shape descriptor and accepting the top retrieval result, we argue that in the case of a large-scale database a more accurate model may be found within a neighborhood of the descriptor. More importantly, we propose that the distinctiveness deficiency of shape descriptors at the instance level can be compensated by a geometry-based re-ranking of its neighborhood. Our approach first leverages the discriminative power of learned representations to distinguish between different categories of models and then uses a novel robust point set distance metric to re-rank the CAD neighborhood, enabling fine-grained retrieval in a large shape database. Evaluation on a real-world dataset shows that our geometry-based re-ranking is a conceptually simple but highly effective method that can lead to a significant improvement in retrieval accuracy compared to the state-of-the-art.

• The paper introduces a dual-stage retrieval method that first ranks CAD model candidates using learned shape descriptors and then refines them with a geometry-based re-ranking using the Modified Single-direction Chamfer Distance (MSCD).
• It demonstrates a notable boost in both Top1 and Top5 retrieval accuracies on the Scan2CAD dataset compared to prior state-of-the-art approaches.
• The robust framework supports applications in augmented reality and robotics by emphasizing the separation of semantic and geometric similarities for precise model retrieval.

Accurate Instance-Level CAD Model Retrieval in a Large-Scale Database

The paper "Accurate Instance-Level CAD Model Retrieval in a Large-Scale Database" proposes a novel method for accurately retrieving CAD models from a vast database based on RGBD scans, with particular attention to instance-level detail. The authors address a significant challenge faced by previous methods, which typically relied on shape descriptors to identify a match within a limited-scale database and often lacked the necessary precision when the database size increased.

Methodology Overview

The authors introduce a two-step retrieval process aimed at overcoming the limitations of existing approaches in large-scale databases. The process begins with a feature-based ranking using a learned shape descriptor to find potential CAD model candidates. These initial candidates are refined through a geometry-based re-ranking step. Crucially, this re-ranking leverages a new robust point set distance metric termed Modified Single-direction Chamfer Distance (MSCD), designed to address the discrepancies between partial, noisy scans and complete CAD models.

Results and Findings

The proposed methodology demonstrates marked improvements in retrieval accuracy over state-of-the-art methods. Quantitative evaluations were conducted on the Scan2CAD dataset, wherein the geometry-based re-ranking exhibited superior performance, with both Top1 and Top5 retrieval accuracy significantly increased. The robust MSCD metric, particularly, was a key factor in reducing error introduced by noise and unobserved sections in RGBD scans.

Significance and Implications

The implications of this research are twofold. Practically, it provides a robust framework for applications in augmented reality and robotics, wherein precise object models are critical for interaction and scene understanding. Theoretically, the work underscores the importance of separating semantic and geometric similarities in large-scale model retrieval tasks, guiding future algorithmic enhancements with a more granular, detail-oriented approach.

Future Directions

While the research presents a considerable step forward, future investigations could further explore the integration of deformation techniques at the retrieval stage to bridge existing discrepancies between database models and actual scanned objects. Additionally, scalability in real-time applications and optimization of computational resources remains a fruitful area for further development.

Overall, this paper presents a comprehensive method addressing the nuanced demands of instance-level retrieval in expansive databases, carving a path for enhanced computer vision applications where detail accuracy is paramount.