MemOcc: Hierarchical Memory for Indoor Continuous Occupancy Mapping
Yang YIRong, Lin Yuxin, Guo Longteng, Song Li, Wang Qunbo, Yu Ming-Ming, Wu Wenjun, Liu Jing
AI summary
Problem
Continuous 3D occupancy mapping in indoor environments struggles with occlusion-induced error propagation and inefficient recomputation when scenes cycle through visibility changes. Existing methods either blindly fuse frames or recompute from scratch, leading to stale errors and poor long-horizon stability.
Approach
MemOcc introduces a plug-and-play memory-augmented framework with a Short-Term Memory Cache for confidence-gated read-write aggregation to stabilize voxel predictions, and a Long-Term Memory Bank for cross-scene prior retrieval and initialization.
Key results
- Reduces error propagation by 25% and improves mapping speed
- Achieves state-of-the-art mIoU with up to 9.76% gain over baselines
- Short-term memory cache boosts mIoU by 9.41% by filtering occlusion noise
- Long-term memory bank enables cross-scene prior reuse and faster initialization
Why it matters
Enables robust, real-time 3D scene understanding for indoor robots and AR systems navigating complex, occluded environments.
Abstract
Indoor 3D occupancy mapping, crucial for robotic perception, struggles with occlusions and reappear- ing surfaces in continuous observations. Existing methods either fuse frames without discernment, causing occlusion- induced errors to persist and contaminate global repre- sentations, or recompute scenes from scratch, sacrificing efficiency and stability. To address these challenges, we pro- pose MemOcc, a novel memory-augmented framework for continuous occupancy mapping using read–write–retrieve operations. MemOcc employs a hierarchical memory de- sign with cooperative short- and long-term tiers. Its Short- Term Memory Cache module uses visibility-gated writes and confidence maps to stabilize voxel predictions and filter occlusion noise, while the Long-Term Memory Bank stores scene priors for rapid retrieval, accelerating convergence in revisited regions. As a plug-and-play module, MemOcc integrates seamlessly with existing 2D-to-3D pipelines with- out altering backbones or training. Experiments on indoor benchmarks demonstrate MemOcc reduces error propaga- tion by 25% and improves mapping speed over state-of- the-art methods, achieving robust, real-time performance. *This research is supported by Artificial Intelligence-National Science and Technology Major Project (2023ZD0121200), the Na- tional Natural Science Foundation of China (62441617, 62437001, 62436001), Beijing Natural Science Foundation (L252146), the Key Research Development Program of Jiangsu Province under Grant BE2023016-3, and the Strategic Priority Research Program of Chinese Academy of Sciences under Grant XDB1350103. 1Beihang University, 2Institute of Automation of the Chinese Academy Sciences, 3Horizon Robotics, 4Beijing Jiaotong Uni- versity, 5University of Chinese Academy of Sciences. *Corre- sponding author. (yirongyang@buaa.edu.cn, longteng.guo@ia.ac.cn, wangqb6@outlook.com) By selectively retaining reliable evidence and enabling efficient retrieval, MemOcc paves the way for scalable indoor perception in robotics and augmented reality.