arXiv:2511.17824v1 Announce Type: new 
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

تم اقتراح نهج جديد يسمى Quality-Aware Loss (QAL) لتحسين مهام إعادة البناء ثلاثية الأبعاد من خلال معالجة قيود أهداف التدريب الحالية مثل Chamfer Distance (CD) وEarth Mover's Distance (EMD) التي تكافح لتحقيق توازن بين الاسترجاع والدقة. يقدم QAL مصطلح جيران الأقرب الموزون بالتغطية ومصطلح جذب نحو الحقيقة الأرضية غير المغطاة، مما يؤدي إلى تحسينات ملحوظة في التغطية عبر قنوات مختلفة.

Se ha propuesto un nuevo enfoque llamado Quality-Aware Loss (QAL) para mejorar las tareas de reconstrucción 3D al abordar las limitaciones de los objetivos de entrenamiento existentes, como la Chamfer Distance (CD) y la Earth Mover's Distance (EMD), que luchan por equilibrar el recuerdo y la precisión. QAL introduce un término de vecino más cercano ponderado por cobertura y un término de atracción hacia la verdad de terreno no cubierta, lo que lleva a mejoras significativas en la cobertura a través de diversas canalizaciones.

Une nouvelle approche appelée Quality-Aware Loss (QAL) a été proposée pour améliorer les tâches de reconstruction 3D en abordant les limites des objectifs d'entraînement existants tels que la Chamfer Distance (CD) et l'Earth Mover's Distance (EMD), qui peinent à équilibrer le rappel et la précision. QAL introduit un terme de voisin le plus proche pondéré par la couverture et un terme d'attraction vers la vérité terrain non couverte, entraînant des améliorations significatives de la couverture à travers divers pipelines.

A new approach called Quality-Aware Loss (QAL) has been proposed to improve 3D reconstruction tasks by addressing the limitations of existing training objectives like Chamfer Distance (CD) and Earth Mover's Distance (EMD), which struggle to balance recall and precision. QAL introduces a coverage-weighted nearest-neighbor term and an uncovered-ground-truth attraction term, leading to significant improvements in coverage across various pipelines.

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

arXiv:2503.08594v3 Announce Type: replace-cross 
Abstract: Autoregressive point cloud generation has long lagged behind diffusion-based approaches in quality. The performance gap stems from the fact that autoregressive models impose an artificial ordering on inherently unordered point sets, forcing shape generation to proceed as a sequence of local predictions. This sequential bias emphasizes short-range continuity but undermines the model's capacity to capture long-range dependencies, hindering its ability to enforce global structural properties such as symmetry, consistent topology, and large-scale geometric regularities. Inspired by the level-of-detail (LOD) principle in shape modeling, we propose PointNSP, a coarse-to-fine generative framework that preserves global shape structure at low resolutions and progressively refines fine-grained geometry at higher scales through a next-scale prediction paradigm. This multi-scale factorization aligns the autoregressive objective with the permutation-invariant nature of point sets, enabling rich intra-scale interactions while avoiding brittle fixed orderings. Experiments on ShapeNet show that PointNSP establishes state-of-the-art (SOTA) generation quality for the first time within the autoregressive paradigm. In addition, it surpasses strong diffusion-based baselines in parameter, training, and inference efficiency. Finally, in dense generation with 8,192 points, PointNSP's advantages become even more pronounced, underscoring its scalability potential.

تم تقديم PointNSP كإطار عمل ذاتي الانحدار لتوليد سحب النقاط ثلاثية الأبعاد، حيث يعالج القيود التي تواجه النماذج التقليدية التي تكافح لالتقاط الاعتماديات على المدى الطويل والخصائص الهيكلية العالمية. تستخدم هذه الطريقة المبتكرة إطارًا توليديًا من الخشن إلى الدقيق يحافظ على الهيكل العام للشكل عند الدقة المنخفضة بينما يقوم بتحسين الهندسة الدقيقة عند المقاييس الأعلى.

PointNSP se ha introducido como un marco autorregresivo para la generación de nubes de puntos 3D, abordando las limitaciones de los modelos tradicionales que luchan por capturar dependencias a largo plazo y propiedades estructurales globales. Este enfoque innovador emplea un marco generativo de grueso a fino que preserva la estructura global de la forma a bajas resoluciones mientras refina la geometría detallada a escalas más altas.

PointNSP a été introduit comme un cadre autorégressif pour la génération de nuages de points 3D, abordant les limitations des modèles traditionnels qui peinent à capturer les dépendances à long terme et les propriétés structurelles globales. Cette approche innovante utilise un cadre génératif de grossier à fin qui préserve la structure globale de la forme à basse résolution tout en affinant la géométrie détaillée à des échelles plus élevées.

PointNSP has been introduced as an autoregressive framework for generating 3D point clouds, addressing the limitations of traditional models that struggle with capturing long-range dependencies and global structural properties. This innovative approach employs a coarse-to-fine generative framework that preserves global shape structure at low resolutions while refining fine-grained geometry at higher scales.

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

Was this article worth reading? Share it

Keywords AI

Dyad

Mapfit