ECCV 2026

Feed-forward Likelihood Maximization for Efficient Indoor Occupancy Prediction

Guangcheng Chen^1,3

Lihuang Fang^1,3

Huaqi Tao^1,3

Yicheng He^1,3

Li He^1,2,*

Hong Zhang^1,3,*

¹Southern University of Science and Technology

²Guangdong Laboratory of Artificial Intelligence and Digital Economy (Shenzhen)

³Shenzhen Key Laboratory of Robotics and Computer Vision

^* Corresponding authors

arXiv Code

Main teaser figure — Performance and Visualization of Inference

TL;DR

We propose a novel framework for occupancy prediction, Feed-forward Likelihood Maximization, that is able to train a neural network to dramatically relocate and deform primitives to model a scene. The resulting method, FLM-Occ, achieves significant performance improvement in both accuracy and efficiency on the Occ-ScanNet dataset.

ABSTRACT

Recent indoor occupancy prediction methods adopt Gaussian primitives as a sparse 3D representation for computational efficiency. However, their training relies on voxel classification, which imposes only local constraints and lacks global supervision on the distribution of the primitives. Therefore, they inevitably predict spurious primitives in empty regions, undermining both representational and computational efficiency. To address this, we propose Feed-forward Likelihood Maximization (FLM), a novel framework that reformulates occupancy prediction as voxel distribution estimation. In FLM, a network is trained to predict a mixture model that maximizes the likelihood over ground-truth occupied voxels in a feed-forward manner. To enable end-to-end training of networks and voxelization of a standard mixture model, we define mixture weights as normalized primitive volumes to implicitly enforce simplex constraints and derive novel voxelization formulas. Based on FLM, we present FLM-Occ, the first method capable of relocating primitives over long distances to model a scene. On Occ-ScanNet, FLM-Occ achieves superior accuracy using only 32 superquadrics, $2.7\%$ of the primitives of prior SoTA, while running $3.7\times$ faster.

KEY CONTRIBUTIONS

Identifying the gradient vanishing issue that prevents previous methods from relocating geometric primitives over long-distance to model a scene.

Feed-forward Likelihood Maximization (FLM), a new framework for indoor occupancy prediction, that allows a neural network to learn substantial relocation and deformation of the primitives.

FLM-Occ, an indoor occupancy prediction method that achieves significant improvements in both accuracy and efficiency.

METHOD

Method Overview

Framework diagram — Overview of our proposed framework

RESULTS

Qualitative Results

Result 1 — Comparisons between SplattSC and FLM-Occ using different primitive configurations

Sample image for scene0024_00 frame 00012 — scene0024_00 / 00012

SplatSSC

FLM-Occ

Sample image for scene0032_00 frame 00076 — scene0032_00 / 00076

SplatSSC

FLM-Occ

Sample image for scene0040_00 frame 00005 — scene0040_00 / 00005

SplatSSC

FLM-Occ

Sample image for scene0070_00 frame 00017 — scene0070_00 / 00017

SplatSSC

FLM-Occ

EVALUATION

Quantitative Results

Method	Primitive	Number	Distance (m)	IoU (%) ↑	mIoU (%) ↓	FPS ↑
EmbodiedOcc	Gaussian	16200	0.08	53.6	45.2	7.8
SplatSSC	Gaussian	1200	0.15	62.8	51.8	8.7
Ours	Gaussian	64	1.40	64.7	56.3	32.0
Ours	Superquadric	32	1.32	65.3	55.9	32.1
Ours	Gaussian	1024	1.25	71.2	63.6	23.6
Ours	Superquadric	1024	1.20	71.7	64.0	25.4

The distance values represent the average moving distances between the initial and final positions of the primitives. More results can be found in our paper.

CITATION

BibTeX

@article{flmocc2026gcchen,
  title={FLM-Occ: Feed-forward Likelihood Maximization for Efficient Indoor Occupancy Prediction},
  author={Chen, Guangcheng and Fang, Lihuang and Tao, Huaqi and He, Yicheng and He, Li and Zhang, Hong},
  journal={arXiv preprint arXiv:},
  year={2026}
}