Hierarchical Incremental Beta Splatting for Online Reconstruction
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Hierarchical Incremental Beta Splatting for Online Reconstruction".
Jane: Here's the gist of the paper. Online 3D Gaussian Splatting reconstructs a scene during capture: each keyframe spawns primitives once and refines them for only a few dozen iterations, leaving little chance to fix poor initialization or missing detail<ref:2611.08288#pg2>.
Tom: Core Methodology The paper introduces Hierarchical Incremental Beta Splatting (HIBS), which brings beta kernels to incremental reconstruction<ref:2611.08288#pg4>.
Jane: This method replaces the Gaussian falloff with a bounded beta kernel, whose shape parameter b interpolates between a soft, Gaussian-like profile and a flat-topped, near-opaque surfel<ref:2611.08288#pg5>.
Tom: The confidence dial is controlled by this shape parameter b, which ranges from soft to solid<ref:2611.08288#pg6>.
Paper summary: Tom: So looking at the full picture of "Hierarchical Incremental Beta Splatting for Online Reconstruction," it seems like they’ve built a system that handles the messy reality of capturing data in real time >
Jane: They take the core idea of beta kernels, which naturally have this bounded profile between soft and solid, and they adapt it for online capture by using geometric evidence to set that confidence dial >
Lu: What I find most important is how they treat the primitive's lifecycle. They don't just spawn and forget; they use error-driven relocation to send dead primitives back into areas where things are poorly explained >
Meng: That recycling of capacity sounds like something you’d actually need when you’re deploying models to run on limited hardware or dealing with noisy input streams at scale >
Tom: And that short optimization burst, followed by fused rasterization, is the mechanism they use to quickly polish those primitives based on that geometric evidence >
Jane: It boils down to encoding a continuous notion of solidity into the primitive's shape field itself, where low beta values mean you have earned that solid status through observation >
Lu: They are using this as an evidence-driven confidence state, and the fact that they can validate their merging procedure with those Monte Carlo estimates is a strong piece of evidence for the overall structure >
Meng: It’s not just about getting one big number; it’s about having these different components—the spawning, the relocation, the merging—all work together to keep things stable during a continuous stream >
Tom: The implication for people using this is that you get faster and higher quality reconstruction throughout the whole capture process without needing massive offline densification steps >
Jane: It means less waiting for a final result and more usable three dee data as the scene is being built, even if you only have a few dozen iterations refining each frame > <ref:2611.08288#pg1>
Conclusion: Tom: So, we’re looking at "Hierarchical Incremental Beta Splatting for Online Reconstruction." Basically, this paper tackles how to build a three dee scene while you're still recording data frame by frame, without having all the time to process everything at once.
Jane: That’s right. The main idea is they use something called beta kernels, which naturally let you control how solid or soft a three dee point is. They adapt that for online capture by making it aware of what the scene looks like so far.
Lu: The authors introduce Hierarchical Incremental Beta Splatting, which uses these beta kernels in an incremental way. Instead of just adding new points, they refine the existing ones a few dozen times per keyframe, which is important because you don't have infinite time to fix errors.
Meng: If I'm looking at this from a production standpoint, what I’m interested in is how they manage the primitive lifecycle when you can’t densify offline. The paper says they use error-driven relocation to send dead primitives back into spots where the fit is bad.
Tom: That recycling of capacity sounds like something that would help keep things from just accumulating garbage data over a long capture session. I'm wondering how much time that actually saves you in terms of GPU cycles.
Jane: Well, they show results on Mip-NeRF three hundred sixty where this method improved the picture quality by up to zero point eight decibels and cut down optimization time by about forty percent when running it on the fly.
Lu: They also talk about a closed-form way to merge these beta primitives for large scenes, which lets them keep track of the shape statistics without having to do all that heavy minimization work every time they combine things.
Meng: A closed-form merge is good, but what I really need to know is if it actually works reliably when you’re dealing with noisy input data from a camera. I’m skeptical about those perfect mathematical merges in real-world scenarios.
Tom: The ablation studies suggest that things like confidence-aware spawning and error relocation all help the final reconstruction quality, which is encouraging because it means each piece of the system is contributing something concrete to the result.
Jane: It seems they’ve essentially created a system where you get a live map of confidence on your scene as you capture it, visualized by this shape field they generate.
Lu: That shape field acts like an immediate confidence map for the reconstruction because low beta values mean the geometry is firmly committed, which is a great way to track progress.
Tom: So, in short, this framework redesigns how primitives live and die during incremental capture to keep things fast and high quality as you go. We need to see if that live confidence map holds up under different kinds of scene complexity.
Linus Franke
University of Wurzburg
cs.CV, cs.GR
Submitted: 2026-11-05
Updated: 2026-11-05
Comments: accepted to WACV 2027, project page: https://linus-franke.com/hibs/
Project page: https://linus-franke.com/hibs
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: Online 3D Gaussian Splatting reconstructs a scene during capture: each keyframe spawns primitives once and refines them for only a few dozen iterations…
Key concepts
- Hierarchical Incremental Beta Splatting (HIBS)
- This method replaces standard Gaussian falloff with a bounded beta kernel. The shape parameter 'b' interpolates between a soft, Gaussian profile and a flat-topped, near-opaque surfel. This allows the system to continuously control the level of solidity, ranging from soft to solid, based on accumulated evidence.
- Confidence Initialization
- The system initializes primitive opacity and shape using geometric cues. It combines geometric depth observability with cross-validation between two independent depth estimators. This results in a confidence parameter 'c' that determines whether a primitive starts as a hard patch or soft probe.
- Error-Driven Relocation
- This mechanism manages the limited optimization time by recycling near-dead primitives into photometric-error hotspots. It reallocates unused capacity to regions where the reconstruction is currently poorly explained, improving detail in areas needing more refinement.
Terminology
Summary
Online 3D Gaussian Splatting reconstructs a scene during capture: each keyframe spawns primitives once and refines them for only a few dozen iterations, leaving little chance to fix poor initialization or missing detail<ref:2611. p. 0 Compared to i3DGS [42], we achieve higher quality with about 20% faster optimization. Abstract
Online 3D Gaussian Splatting reconstructs a scene during
capture: each keyframe spawns primitives once and refines
them for only a few dozen iterations, leaving little chance to
fix poor initialization or missing detail.Where the paper says it
Core Methodology
The paper introduces Hierarchical Incremental Beta Splatting (HIBS), which brings beta kernels to incremental reconstruction<ref:2611.08288#pg4>. This method replaces the Gaussian falloff with a bounded beta kernel, whose shape parameter b interpolates between a soft, Gaussian-like profile and a flat-topped, near-opaque surfel<ref:2611. p. 1 We instead use the beta kernel [35], whose bounded profile has a learnable shape parameter b that continuously
interpolates between a Gaussian-like volume and a flattopped, near-opaque surfel. A single beta primitive can thus
be a flat-topped surface patch while retaining a soft representation for uncertain geometry, thin structures, or volume. Their shape parameter interpolates between a
soft, Gaussian-like profile and a flat-topped, near-opaque
surfel, acting as a confidence dial for geometry. Our
confidence-aware spawning uses geometric observability
and multi-view depth agreement to initialize both opacity
and shape, producing solid primitives where geometry is
verifiable and soft, correctable primitives where it is not.Where the paper says it
Where the paper says it
Confidence Initialization
The authors employ confidence-aware spawning, initializing opacity and shape from independent geometric cues<ref:2611.08288#pg5>. This involves combining geometric depth observability with cross-validation between two independent depth estimators<ref:2611. p. 3 With the short optimization burst following and no guarantee of repeated observations, opacity and shape must be initialized well from
observable geometry. We combine geometric depth observability with cross-validation between two independent
depth estimators, as also depicted in Fig. Spawning/parameter initialization. Based on depth
observability cz and agreement of depths (orange window) cmvs,
the beta primitives are initialized as hard patches or soft probes. Spawning/parameter initialization. Based on depth
observability cz and agreement of depths (orange window) cmvs,
the beta primitives are initialized as hard patches or soft probes.Where the paper says it
Where the paper says it
Where the paper says it
Primitive Lifecycle Management
The primitive lifecycle is managed through several mechanisms to compensate for the lack of offline densification<ref:2611.08288#pg6>. These include error-driven relocation, which recycles near-dead primitives into photometric-error hotspots<ref:2611. p. 1 • We introduce confidence-aware spawning, which initializes opacity and shape from independent depth agreement
and geometric observability. • We introduce error-driven relocation, a render-preserving
online alternative to clone/split densification that recycles
near-dead primitives into photometric-error hotspots. Confidence-aware spawning combines two independent geometric cues, Multi-view stereo (MVS) agreement with an aligned monocular prior and depth observability, to initialize each primitive along this soft-to-solid
axis: well-constrained candidates are spawned solid, while
ambiguous ones remain soft and correctable. As additional
views arrive, a lifetime controller updates the same parameter from accumulated multi-view evidence, progressively
turning confirmed geometric confidence into solidity. The beta representation is also performant. Its bounded
support yields an exact tile-culling bound and faster rasterization, while its closed-form moments enable a shapeaware level-of-detail hierarchy for large captures.Where the paper says it
Where the paper says it
Where the paper says it
Optimization and Merging
The system utilizes a short optimization burst, typically involving 30 iterations, followed by fused rasterization<ref:2611.08288#pg6>. To manage large scenes, a closed-form merge for beta primitives is derived using moment matching instead of KL minimization<ref:2611. p. 2 Near-dead primitives are relocated
to high-error regions, followed by a short 30-iteration optimization burst with our fused rasterizer. Periodically, small-footprint primitives
are merged hierarchically using closed-form moment matching that preserves their solidity. As
online densification, error-driven relocation recycles stale
primitives into photometric-error hotspots. Bounded beta
kernel support enables exact tile culling and faster rasterization, and a closed-form, shape-preserving merging formulation provides a level-of-detail (LOD) hierarchy for fast
optimization and real-time rendering.Where the paper says it
Where the paper says it
Performance and Results
Experiments demonstrate that HIBS improves PSNR by up to 0.8 dB while reducing optimization time by up to 40% on Mip-NeRF 360 in the on-the-fly protocol<ref:2611. p. 0 On Mip-NeRF 360 in
the on-the-fly protocol, HIBS improves PSNR by up to 0. 8
dB while reducing optimization time by up to 40%. 1 evaluates on the MipNeRF360 (otf), TUM, and
StaticHikes scenes against pose-free and online baselines. In this benchmark, our method outperforms our closest
competitor i3DGS (except time on TUM), and is consistently among the strongest methods overall. Quality During Reconstruction
An online system should improve throughout reconstruction, not only in the final model. 7 (top) reports mean
PSNR over the test views registered so far as reconstruction
progresses.Where the paper says it
Where the paper says it
Where the paper says it
Ablation and Sensitivity
Ablation studies confirm that components like confidence-aware spawning, solidification, and error relocation all improve reconstruction quality<ref:2611.08288#pg8>. The shape parameter b acts as an evidence-driven confidence state, where low β denotes earned geometric commitment<ref:2611. p. 4 Thus β acts as an evidence-driven confidence state: low
β denotes earned geometric commitment, while high β denotes unresolved or view-dependent evidence. The resulting shape field provides an immediate confidence map of
the reconstruction, as illustrated in Fig. Building on beta kernels, we use kernel shape as a confidence dial: each keyframe spawns primitives with opacity and shape initialized
from measured multi-view confidence. Short optimization bursts, fused rasterization, error-driven relocation, and closed-form LoD merging
enable efficient reconstruction. First, the mass Z, covariance Cov = ξ(β)Σ with shrink
factor ξ(β) = 4. 5), and Mardia kurtosis κ(β) =
15(β + 2. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. Together, these components yield an online reconstruction system in which the primitive is adapted to the short
optimization horizon rather than relying on offline densification to compensate for it. Our experiments show improved reconstruction quality and efficiency across online,
benchmark, and large-scale captures. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The authors conclude by presenting an online reconstruction framework that redesigns the primitive lifecycle around the short optimization horizon of incremental capture<ref:2611.08288#pg8>. The bounded beta kernel encodes a continuous notion of solidity, observability-aware spawning initializes this confidence from geometric evidence, and error-driven relocation reallocates unused capacity to poorly explained regions<ref:2611. p. 7 We present an online reconstruction framework that redesigns the primitive lifecycle around the
short optimization horizon of incremental capture. Our
bounded beta kernel encodes a continuous notion of solidity, observability-aware spawning initializes this confidence
from geometric evidence, and error-driven relocation reallocates unused capacity to poorly explained regions. Closedform moment matching enables a shape-aware hierarchy,
while bounded-support rasterization and fused optimization make the representation efficient for large-scale online reconstruction. Together, these components yield faster
and higher-quality reconstruction throughout capture, while
turning the primitive shape field into a live confidence visualization. Thus β acts as an evidence-driven confidence state: low
β denotes earned geometric commitment, while high β denotes unresolved or view-dependent evidence. The resulting shape field provides an immediate confidence map of
the reconstruction, as illustrated in Fig. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. • We develop a fast beta rasterizer with a shape-correct
culling bound that exploits the kernel’s compact support. The project page and source code can be found at:
https://linus-franke. Hierarchical Incremental Beta Splatting for Online Reconstruction
Linus Franke
University of Wurzburg ¨
https://linus-franke. com/hibs
input stream online reconstruction
β primitives ours LPIPS: 0. Hierarchical Incremental Beta Splatting for Online Reconstruction
Linus Franke
University of Wurzburg ¨
https://linus-franke. com/hibs
input stream online reconstruction
β primitives ours LPIPS: 0. Closedform moment matching enables a shape-aware hierarchy,
while bounded-support rasterization and fused optimization make the representation efficient for large-scale online reconstruction. Together, these components yield faster
and higher-quality reconstruction throughout capture, while
turning the primitive shape field into a live confidence visualization. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. 1
arXiv:2611. 08288v1 [cs. 1
arXiv:2611. 08288v1 [cs. 08288v1 [cs. CV] 5 Nov 2026 08288v1 [cs. CV] 5 Nov 2026 Component ablation (leave-one-out from full method). Top table: Metrics are averaged over all nine OtF scenes (Tab. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. The resulting shape field provides an immediate confidence map of
the reconstruction, as illustrated in Fig. Error-Driven Incremental Relocation
The overall amount of new detail that can be added per
keyframe is limited, and offline clone/split heuristics require longer optimization. Feed-forward and
hybrid methods [23, 30, 31, 33] trade scene-specific optimization for speed, typically at reduced fidelity. Closest
to our setting, On-the-Fly NVS [41] and i3DGS [42] reconstruct image streams with short optimization bursts and
no densification, using learned matching and hierarchical
structures. Conversely, a
low photometric error does not guarantee correct geometry. When the triangulation baseline is small relative to depth,
many depths explain the observed views similarly well, yet
only a narrow range remains consistent under parallax. A
Gaussian is fully opaque only at its center and translucent
elsewhere; representing a solid surface therefore requires
overlapping many Gaussians, together with the densification and long optimization that offline 3DGS can afford. 1
arXiv:2611. Against exhaustiveGLOMAP references (Tab. 16), we remain competitive at
7k iterations, while 30k iterations achieve higher quality at
∼ 6–12× our runtime and require the complete capture to
be available beforehand. Against exhaustiveGLOMAP references (Tab. 16), we remain competitive at
7k iterations, while 30k iterations achieve higher quality at
∼ 6–12× our runtime and require the complete capture to
be available beforehand. • We develop a fast beta rasterizer with a shape-correct
culling bound that exploits the kernel’s compact support. The project page and source code can be found at:
https://linus-franke.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive Primitive Representation: The system can utilize a bounded beta kernel whose shape parameter b acts as a confidence dial, interpolating between a soft Gaussian-like profile and a flat-topped, near-opaque surface patch based on measurable multi-view confidence. This allows the system to represent geometry that is either verifiable (solid) or correctable (soft).
-
Confidence-Aware Initialization: The AI system can initialize primitives using geometric observability and multi-view depth agreement to set opacity and shape, producing solid primitives where geometry is verifiable and soft, correctable primitives where it is not. This initialization uses a confidence measure c calculated as a product of depth observability (cz) and cross-validation between two independent depth estimators (cmvs).
-
Error Correction Mechanism: The system can employ error-driven relocation to recycle near-dead primitives into photometric error hotspots, which provides
additional capacity precisely where the current model is deficient.
This mechanism serves as an online alternative to clone/split densification. -
Shape Control via Confidence Field: The AI system can maintain a live confidence visualization through a shape field where the shape parameter b is updated based on exponential moving averages of photometric error, acting as an
evidence-driven confidence state.
Low beta values denote earned geometric commitment, while high beta values denote unresolved or view-dependent evidence. -
Efficient Merging for Large Scenes: The AI system can perform closed-form moment matching to merge primitives hierarchically. This process preserves the shape statistics and enables a level-of-detail hierarchy for fast optimization and real-time rendering, avoiding issues with KL minimization that are unsuitable for compact kernels.
-
Accelerated Optimization: The system can use a faster beta rasterizer with a shape-correct culling bound derived from the kernel's compact support to ensure both correctness and efficiency during short optimization bursts, reducing computation time significantly compared to standard Gaussian cutoffs.
Abstract
Online 3D Gaussian Splatting reconstructs a scene during capture: each keyframe spawns primitives once and refines them for only a few dozen iterations, leaving little chance to fix poor initialization or missing detail. Gaussians are also inherently soft, needing many overlaps for solid surfaces. We present Hierarchical Incremental Beta Splatting (HIBS), which brings beta kernels to incremental reconstruction. Their shape parameter interpolates between a soft, Gaussian-like profile and a flat-topped, near-opaque surfel, acting as a confidence dial for geometry. Our confidence-aware spawning uses geometric observability and multi-view depth agreement to initialize both opacity and shape, producing solid primitives where geometry is verifiable and soft, correctable primitives where it is not. As online densification, error-driven relocation recycles stale primitives into photometric-error hotspots. Bounded beta kernel support enables exact tile culling and faster rasterization, and a closed-form, shape-preserving merging formulation provides a level-of-detail (LOD) hierarchy for fast optimization and real-time rendering. On Mip-NeRF 360 in the on-the-fly protocol, HIBS improves PSNR by up to 0.8 dB while reducing optimization time by up to 40%. Project page: https://linus-franke.com/hibs/
Sources
- DISTWAR: Fast Differentiable Rendering on Raster-based Rendering Pipelines
- DROID-Splat: Combining end-to-end SLAM with 3D Gaussian Splatting
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs