Sparse-Mono: Sparse Gaussian Splatting for Monocular Simultaneous Localization and Mapping

Abstract Feed-forward Gaussian simultaneous localization and mapping (SLAM) can reduce scene-specific optimization by directly predicting Gaussian primitives, but many existing feed-forward methods still generate a dense set of Gaussians for each mapping keyframe, even when much of the current view is already represented in the persistent map. This causes redundant prediction and unnecessary map growth during continuous mapping. We propose Sparse-Mono, a persistent-map-conditioned neural framework for sparse incremental Gaussian generation in monocular SLAM. Instead of reconstructing each keyframe independently, Sparse-Mono jointly encodes the current observation, the rendered persistent map, and their photometric and geometric discrepancies. A multi-scale network with spatial Transformer adapters and a causal event-state Transformer produces map-conditioned features from the current observation, the persistent map, and preceding mapping context. Separate location and count heads then determine where new Gaussians are needed and how many should be generated, after which complete geometry and appearance are decoded only at the selected locations. Valid frontend depths are locally refined, while missing or invalid depths are predicted relative to the current event scale. Experiments on Replica, BundleFusion, and KITTI Odometry show that Sparse-Mono achieves higher average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) with substantially more compact maps. On Replica, it improves PSNR by 1.05 dB over MonoGS and uses 89.3% fewer Gaussians than VINGS-Mono, while maintaining comparable online processing latency. These results demonstrate that conditioning Gaussian generation on the existing map provides an effective way to control map growth while maintaining strong rendering quality.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11445289/v1
Primary Topic
Robotics and Sensor-Based Localization
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Sparse-Mono: Sparse Gaussian Splatting for Monocular Simultaneous Localization and Mapping

Fanliang Bu, Jianpeng Xu, Yifan Wang
Research Square
Robotics and Sensor-Based Localization
preprint

Sparse-Mono: Sparse Gaussian Splatting for Monocular Simultaneous Localization and Mapping

Fanliang Bu, Jianpeng Xu, Yifan Wang
preprint en

Abstract

Abstract Feed-forward Gaussian simultaneous localization and mapping (SLAM) can reduce scene-specific optimization by directly predicting Gaussian primitives, but many existing feed-forward methods still generate a dense set of Gaussians for each mapping keyframe, even when much of the current view is already represented in the persistent map. This causes redundant prediction and unnecessary map growth during continuous mapping. We propose Sparse-Mono, a persistent-map-conditioned neural framework for sparse incremental Gaussian generation in monocular SLAM. Instead of reconstructing each keyframe independently, Sparse-Mono jointly encodes the current observation, the rendered persistent map, and their photometric and geometric discrepancies. A multi-scale network with spatial Transformer adapters and a causal event-state Transformer produces map-conditioned features from the current observation, the persistent map, and preceding mapping context. Separate location and count heads then determine where new Gaussians are needed and how many should be generated, after which complete geometry and appearance are decoded only at the selected locations. Valid frontend depths are locally refined, while missing or invalid depths are predicted relative to the current event scale. Experiments on Replica, BundleFusion, and KITTI Odometry show that Sparse-Mono achieves higher average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) with substantially more compact maps. On Replica, it improves PSNR by 1.05 dB over MonoGS and uses 89.3% fewer Gaussians than VINGS-Mono, while maintaining comparable online processing latency. These results demonstrate that conditioning Gaussian generation on the existing map provides an effective way to control map growth while maintaining strong rendering quality.

Research Square
China People's Public Security University (CN)
Robotics and Sensor-Based Localization
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon