Monocular Gaussian SLAM stops painting every keyframe: Sparse-Mono reads the existing map first and uses 89.3% fewer Gaussians than VINGS-Mono on Replica
Synopsis
Monocular Gaussian SLAM predicts a full set of Gaussians for every mapping keyframe, so regions already represented get re-predicted and the map keeps growing. Sparse-Mono encodes the current observation, the persistent map rendered from the current viewpoint, and their photometric and geometric discrepancies, then decides where additions are needed and how many before decoding complete Gaussians only at the selected locations. On Replica it reaches an average PSNR 1.05 dB above MonoGS and uses 89.3% fewer Gaussians than VINGS-Mono, at comparable online processing latency.
Interpretation
The mapping unit in monocular Gaussian SLAM shifts from generating a full frame of Gaussians per keyframe to filling only the gaps relative to the existing persistent map, so map growth follows the difference between the current observation and what is already represented. Earlier feed-forward methods such as Flash-Mono directly predict camera states and per-pixel Gaussian attributes, and still generate a full set of Gaussians for each mapping keyframe even when much of the current view is already represented in the persistent map. Across eight Replica scenes, Sparse-Mono averages 27.45 dB PSNR against 26.40 dB for MonoGS, a gain of 1.05 dB, and ends with 88,450 Gaussians versus 825,906 for VINGS-Mono, 89.3% fewer.
Each mapping event is split into three views: what is currently observed, what the map already explains, and where the two disagree, and these drive the choice of addition locations and count. The three input streams are the observation, rendered-map and discrepancy streams with 6, 9 and 5 channels; the discrepancy stream carries absolute RGB residuals, normalized log-depth residuals clipped to [-10,10], and a joint depth-validity mask. On Replica, the observation stream alone gives 23.30 dB PSNR, adding the rendered-map stream raises it to 27.04 dB, and adding the discrepancy stream reaches 27.45 dB.
How many Gaussians to add and where to put them come from two separate prediction heads, and complete geometry and appearance are decoded only at the selected locations. The location head proposes eight lightweight candidate slots per quarter-resolution cell, each with an addition score, a subcell offset and an insertion type; the count head regresses the event's addition count from globally pooled features plus a five-dimensional event descriptor. On Replica, keeping the predicted count but ranking by rules gives 26.02 dB PSNR and 99,042 Gaussians, keeping learned ranking but fixing allocation at 3% of candidates gives 26.87 dB and 116,503 Gaussians, and using both learned components gives 27.45 dB and 88,450 Gaussians.
Perspective
Evaluation covers eight Replica indoor scenes, five BundleFusion scenes and KITTI Odometry sequences 00, 05, 06 and 07 at 320×240 rendering resolution, all run online with fixed network weights on an RTX 4070 SUPER and without post-sequence offline refinement.
The LPIPS advantage is not consistent, and Depth L1 error alone does not establish full surface-geometric fidelity; mapping still depends on frontend pose and depth estimates, and the 300.3-million-parameter network still incurs storage and inference costs. Backend refinement iteration budgets differ across methods, so latency differences cannot be attributed entirely to sparse allocation.