QAL: A Loss for Recall Precision Balance in 3D Reconstruction

2025-11-21

AI Summary

Existing 3D reconstruction loss functions—e.g., Chamfer Distance—struggle to jointly optimize recall and precision, leading to geometric distortions in thin structures and sparse regions. To address this, we propose the Quality-Aware Loss (QAL), the first loss that explicitly decouples coverage (recall) and attraction (precision): it introduces a ground-truth point coverage-weighted term for recall and an attraction term for uncovered ground-truth points to enforce precision. QAL requires no architectural modifications and is plug-and-play for point cloud completion and generation tasks (e.g., PCN, ShapeNet). Extensive experiments across multiple datasets and backbone networks demonstrate that QAL consistently reduces Chamfer Distance by an average of 4.3 points and improves state-of-the-art alternative metrics by 2.8 points. Furthermore, grasp evaluation on GraspNet confirms significantly enhanced reconstruction fidelity, directly improving robotic manipulation reliability.

Technology Category

• Computer Vision: 3D Computer Vision
• Machine Learning: Calibration & Uncertainty Quantification
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Volumetric learning underpins many 3D vision tasks such as completion, reconstruction, and mesh generation, yet training objectives still rely on Chamfer Distance (CD) or Earth Mover's Distance (EMD), which fail to balance recall and precision. We propose Quality-Aware Loss (QAL), a drop-in replacement for CD/EMD that combines a coverage-weighted nearest-neighbor term with an uncovered-ground-truth attraction term, explicitly decoupling recall and precision into tunable components. Across diverse pipelines, QAL achieves consistent coverage gains, improving by an average of +4.3 pts over CD and +2.8 pts over the best alternatives. Though modest in percentage, these improvements reliably recover thin structures and under-represented regions that CD/EMD overlook. Extensive ablations confirm stable performance across hyperparameters and across output resolutions, while full retraining on PCN and ShapeNet demonstrates generalization across datasets and backbones. Moreover, QAL-trained completions yield higher grasp scores under GraspNet evaluation, showing that improved coverage translates directly into more reliable robotic manipulation. QAL thus offers a principled, interpretable, and practical objective for robust 3D vision and safety-critical robotics pipelines

Problem

• Balancing recall and precision in 3D reconstruction volumetric learning
• Replacing Chamfer Distance limitations with tunable coverage components
• Improving thin structure recovery for robotic manipulation reliability

Innovation

• Quality-Aware Loss replaces Chamfer and Earth Mover distances
• It decouples recall and precision into tunable components
• Combines coverage-weighted and uncovered-ground-truth attraction terms