CVPR 2026

SAMosaic3D: Modular Scene Assembly for
Real-Time 3D Segment Anything

School of Information, Renmin University of China
* Corresponding author
SAMosaic3D teaser

(a) Offline methods require full scene reconstruction, unsuitable for real-time perception. (b) Prior online methods treat SAM masks as final instance units, causing spatial fragments and wrong merges. (c) SAMosaic3D reinterprets masks as learnable tiles and assembles them into complete, consistent 3D instances.

Abstract

Online 3D instance segmentation is a critical capability for embodied agents navigating in dynamic environments. However, a fundamental challenge remains in adapting powerful 2D foundation models, like SAM, to 3D online segmentation. Naively lifting SAM's 2D masks to 3D results in severe spatial fragmentation, where a single object is shattered into multiple disconnected parts, especially under occlusion. Subsequent attempts to link these fragments over time via conventional 3D IoU-based tracking prove highly fragile: they struggle to handle occlusions or topological changes, ultimately causing catastrophic identity drift. Departing from such post-processing approaches, we reframe online segmentation as a learnable composition problem. We introduce SAMosaic3D, a differentiable framework that treats SAM-derived masks as "mosaic tiles" and learns to assemble them into temporally consistent 3D instances. SAMosaic3D comprises two key components: Fragment-to-Instance Adaptive Assembly that aggregates fragments through soft-gated attention, and Instance-to-Scene Online Merging that employs cascaded semantic-geometric matching to preserve object identities—replacing rigid IoU thresholds with learnable association guided by observation maturity. Evaluations on ScanNet, ScanNet200, SceneNN and 3RScan datasets demonstrate state-of-the-art performance and zero-shot cross-dataset generalization.

Method Overview

SAMosaic3D framework

SAMosaic3D decomposes online open-vocabulary 3D segmentation into two learnable stages: Fragment-to-Instance Adaptive Assembly groups SAM-derived fragments into coherent instance queries, and Instance-to-Scene Online Merging associates instances with a persistent scene memory across frames. Both stages are jointly trainable end-to-end with hierarchical supervision.

Fragment-to-Instance Assembly

Scene-adaptive grouping predicts instance count and generates a differentiable co-instance affinity matrix. Soft-gated attention enables intra-instance fragments to aggregate complementary information while suppressing inter-instance interference.

Fragment-to-Instance module

Instance-to-Scene Merging

A two-tier memory partitions scene context into short-term and long-term tiers. Current instances merge via learnable cross-attention for incomplete observations, with geometric verification against long-term memory using 3D IoU.

Instance-to-Scene module

Experimental Results

Within-dataset evaluation on ScanNet and SceneNN validation sets.

Method ScanNet SceneNN
APAP50AP25 APAP50AP25
Offline methods
TD3D46.271.181.3------
OneFormer3D59.378.886.7------
Online methods
INS-Conv--57.4----57.6--
TD3D-MA39.060.571.326.042.859.2
ESAM-E†41.660.175.627.548.764.6
ESAM-E+FF†‡42.661.977.133.353.662.5
AutoSeg3D†43.462.577.433.152.663.8
SAMosaic3D (Ours)45.365.978.533.256.569.3
SAMosaic3D† (Ours)46.567.780.135.158.271.8

Class-agnostic 3D instance segmentation on ScanNet200 validation set.

MethodVFMAPAP50AP25FPS
Offline methods
SAMPro3DSAM18.032.856.1--
SAI3DSemanticSAM28.247.267.9--
Online methods
SAM3DSAM20.235.755.50.4
ESAMSAM42.263.779.60.7
ESAM-EFastSAM43.465.480.910.6
AutoSeg3DSAM45.566.781.00.7
AutoSeg3D-EFastSAM46.267.981.710.1
SAMosaic3D (Ours)SAM46.168.584.20.7
SAMosaic3D (Ours)FastSAM48.769.385.411.2

Zero-shot cross-dataset generalization from ScanNet200 (no fine-tuning on target domains).

Method ScanNet200 → SceneNN ScanNet200 → 3RScan
APAP50AP25 APAP50AP25
Offline methods
SAMPro3D12.625.853.23.98.021.0
SAI3D18.634.765.78.116.937.0
Online methods
SAM3D15.130.051.86.213.033.9
ESAM28.852.269.314.131.259.6
ESAM-E†28.650.471.013.929.458.8
AutoSeg3D29.753.671.916.032.460.7
AutoSeg3D-E†30.254.172.816.834.361.0
SAMosaic3D (Ours)31.756.874.216.334.763.1
SAMosaic3D† (Ours)32.457.874.717.035.563.7

Qualitative Results

Qualitative comparison

Qualitative comparison of 3D instance segmentation on challenging indoor scenes from SceneNN dataset. SAMosaic3D eliminates over-segmentation artifacts through differentiable soft assignment and co-instance affinity modeling.

Citation

@inproceedings{wang2026samosaic3d,
  title     = {SAMosaic3D: Modular Scene Assembly for Real-Time 3D Segment Anything},
  author    = {Wang, Peng and Wang, Yongcai and Chen, Wang and Cao, Hualong and Yang, Kang and Li, Chunxu and Jie, Wen and Li, Deying},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}