Researchers crack 3D indoor scene AI with 11.38% accuracy boost

Researchers from Qi Zheng’s team have introduced Group-UFD Occ, a groundbreaking method for 3D semantic occupancy prediction in indoor environments. Unlike outdoor scenes, indoor spaces suffer from severe long-tailed data distributions, where rare object categories dominate performance bottlenecks. The team tackled this challenge with a two-pronged approach: a fine-grained semantic grouping strategy and multi-scale, parallel 'main-expert' prediction heads designed to efficiently learn features from tail-class objects through deep regularization.

The optimization layer introduces the Unified Focal-Dice (UFD) loss, a synergistic loss function that dynamically focuses on hard samples at the voxel level while optimizing geometric integrity from a region-based perspective. In experiments on the large-scale EmbodiedScan dataset, the method achieved a relative accuracy improvement of 11.38% over baseline models, with significant gains in several critical long-tailed categories—demonstrating its potential to revolutionize indoor scene understanding for robotics and autonomous systems.