Managing Multi-View Primitive Redundancy in 3D Reconstruction

22 concepts19 papers1 proposal

In 3D computer vision, more data is generally expected to yield better results. However, a recurring paradox in feed-forward reconstruction is that increasing the number of input views can actually degrade output quality. This occurs because most models treat each camera view as an independent source of geometry, leading to primitive redundancy—a phenomenon where the same physical point in space is represented by multiple, uncoordinated geometric primitives.

The Multi-View Paradox and Primitive Stacking

When a model reconstructs a scene from multiple perspectives, it often generates a set of 3D primitives (such as Gaussians, points, or line segments) for every image. Without a mechanism to coordinate these predictions, the resulting scene becomes cluttered with overlapping geometry. Shengjun Zhang et al. describe this as "primitive stacking," where redundant representations are layered on top of each other, leading to visual artifacts and significant memory overhead [Gaussian Graph Network].

This redundancy is not merely a storage issue; it actively harms the visual fidelity of the reconstruction. In autonomous driving contexts, where multiple sensors have overlapping fields of view, Hao LU notes that these redundant points cause "ghosting" artifacts, where the overlapping data from different sensors clashes rather than converges [DrivingRecon]. Similarly, David Charatan et al. observe that simple aggregation of view-dependent primitives leads to inefficient representations that struggle to scale [pixelSplat].

Mechanisms of Redundancy

Redundancy manifests in several ways across different 3D representations:

  1. Geometric Overlap: In feed-forward 3D Gaussian Splatting, independent predictions for the same spatial location result in "floaters" and excessive memory usage [FreeSplat].
  2. Pose Redundancy: Standard models often fail to recognize that the same geometric feature (like a chair leg) is identical even when viewed from different angles or positions. Xin Hu et al. argue that this forces models to waste capacity learning the same shape multiple times [Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation].
  3. Computational Duplication: When using view-dependent structures like cost frustums, overlapping regions are computed multiple times, decreasing efficiency as the number of input images grows [C2F2NeUS].

Strategies for De-duplication and Fusion

To overcome these hurdles, researchers have moved toward unified, view-agnostic representations that fuse information across perspectives before generating geometry.

  • Graph-Based Message Passing: The Gaussian Graph Network (GGN) treats 3D Gaussians as nodes in a graph, using pooling layers to eliminate redundant representations. This allows the model to use up to 10x fewer Gaussians while improving quality [Gaussian Graph Network].
  • Canonical Refinement: Instead of just stacking points, Yiming Wang et al. propose a "Fuse-and-Refine" module that merges overlapping primitives into a canonical 3D space, ensuring cross-view integration [Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting].
  • Information Bottlenecks: To handle large-scale data, Weijie Wang et al. apply the Information Bottleneck principle to compress multi-view inputs into a compact latent state, preventing the "representation overload" that typically causes image blurring in dense-view scenarios [ZPressor].
  • Adaptive Pruning: Models like PixelGaussian dynamically split Gaussians in complex areas and prune them in redundant ones, ensuring that increasing input views enhances the scene rather than cluttering it [PixelGaussian].

Redundancy in 3D Reasoning

The challenge of redundancy extends beyond reconstruction to 3D understanding. When 2D Vision-Language Models (VLMs) process 3D scenes via multiple images, they often encounter the same objects repeatedly, wasting valuable input tokens. Hsiang-Wei Huang et al. developed Dynamic Token Compression (DTC) to merge these redundant visual observations into a consistent spatial representation, reducing token counts by over 90% without losing reasoning performance [Zero-shot 3D Question Answering via Voxel-based Dynamic Token Compression].

By shifting from simple view aggregation to intelligent fusion and pruning, these methods ensure that every primitive in a 3D model provides unique, non-redundant information about the scene.

Key Foundational Work

Learning Attentive and Hierarchical Representations for 3D Shape Recognition (2020) — contributed 1 concept elements to this direction.
Removing Redundant Line Segments of 3D Model for Camera Pose Estimation (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_primitive_redundancy_multi_view_primitiv_84ce76c7,
  title = {primitive redundancy (multi-view primitive redundancy)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/primitive-redundancy-multi-view-primitive-redundancy/rdr_84ce76c7e8b74f2c84ad664e948dcd03},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_primitive_redundancy_multi_view_primitiv_84ce76c7_v1,
  title = {primitive redundancy (multi-view primitive redundancy)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/primitive-redundancy-multi-view-primitive-redundancy/rdr_84ce76c7e8b74f2c84ad664e948dcd03/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake