Group-wise Supervision with Focal-Dice Loss for Long-Tailed Indoor Semantic Occupancy Prediction

2026-07-30

AI Summary

This work addresses the challenge of limited performance on tail classes in indoor scene semantic occupancy prediction, caused by the long-tailed category distribution. To mitigate this issue, the authors propose a hierarchical semantic grouping supervision scheme combined with a multi-scale parallel “backbone-expert” prediction head architecture. A unified Focal-Dice loss function is introduced to jointly optimize voxel-level hard example focusing and region-level geometric completeness. By leveraging fine-grained semantic grouping and a joint optimization strategy, the method significantly enhances learning efficiency for tail classes. Evaluated on the EmbodiedScan dataset, the approach achieves an 11.38% improvement in mean Intersection over Union (mIoU) over the baseline and demonstrates substantial gains in accuracy across multiple critical tail categories.

Abstract

Recently, 3D semantic occupancy prediction has garnered increasing attention for understanding the indoor scene. However, unlike structured outdoor environments, indoor scenes feature a high diversity of object categories that exhibit a severe long-tailed distribution, which has become a core bottleneck limiting the performance of existing models. To tackle this challenge, we propose a novel method, Group-UFD Occ, based on hierarchical semantic supervision and synergistic loss optimization. At the architectural level, we introduce a fine-grained semantic grouping strategy and design multi-scale, parallel ``main-expert'' prediction heads to guide the model in efficiently learning tail-class features through deep regularization. At the optimization level, we introduce the Unified Focal-Dice (UFD) loss. This synergistic loss function dynamically focuses on hard samples at the per-voxel level. Meanwhile, it simultaneously optimizes the geometric integrity of predicted objects from a region-based perspective. We conducted experiments on the large-scale EmbodiedScan dataset. The results demonstrate that our method yields a relative improvement of 11.38\% over the baseline, with substantial accuracy gains in several critical long-tailed categories.

Problem

• long-tailed distribution
• semantic occupancy prediction
• indoor scene understanding
• 3D semantic segmentation
• class imbalance

Innovation

• semantic occupancy prediction
• long-tailed distribution
• group-wise supervision
• Focal-Dice loss
• multi-scale expert heads