← 研踪 PaperTrailarXiv:2611.20533 · v1
RemeshingAgent: An Iterative Agent for Repair-Based Decimation of Generated Building Meshes
首次提交 2026-11-16 13:08
版本更新 2026-11-16 13:08
雷达观测 2026-11-17 13:07
摘要
Aggressively simplifying a generated high-poly building mesh to 0.5-1% of its faces confronts a hard limit. The surface looks complete, yet the mesh is topologically shattered: 10^3-10^4 connected components and non-manifold thin walls. Below a certain budget, every practical decimator "melts": it deletes parts, welds thin walls, or breaks the error tolerance. We attribute this to instance-specific pathology: which components are significant and where the cliff sits vary from mesh to mesh, so no single fixed algorithm suffices, and decimation becomes an iterative loop of cleaning, simplifying, and verifying. We present RemeshingAgent, an agent that turns this loop into a validated process. The melting follows from a triangle-budget decomposition into curvature, crease, component, and genus terms: below the component floor, no output satisfies the repair contract, and cleaning first is the theory-mandated response. At a 0.5% budget the agent stays topology-clean where QEM and vertex clustering melt, matches a commercial remesher on 20 unseen city buildings, and on the 24 training buildings a nine-participant single-workstation blind study prefers our outputs 168:48 (p = 0.0195, subject-level sign test). The trained skill deploys as retrieval replay: 11.7 seconds per mesh on CPU, with no LLM calls.
作者备注:11 pages, 12 figures, 3 tables
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-16 13:08 · 最新观测版本
RemeshingAgent: An Iterative Agent for Repair-Based Decimation of Generated Building Meshes
Aggressively simplifying a generated high-poly building mesh to 0.5-1% of its faces confronts a hard limit. The surface looks complete, yet the mesh is topologically shattered: 10^3-10^4 connected components and non-manifold thin walls. Below a certain budget, every practical decimator "melts": it deletes parts, welds thin walls, or breaks the error tolerance. We attribute this to instance-specific pathology: which components are significant and where the cliff sits vary from mesh to mesh, so no single fixed algorithm suffices, and decimation becomes an iterative loop of cleaning, simplifying, and verifying. We present RemeshingAgent, an agent that turns this loop into a validated process. The melting follows from a triangle-budget decomposition into curvature, crease, component, and genus terms: below the component floor, no output satisfies the repair contract, and cleaning first is the theory-mandated response. At a 0.5% budget the agent stays topology-clean where QEM and vertex clustering melt, matches a commercial remesher on 20 unseen city buildings, and on the 24 training buildings a nine-participant single-workstation blind study prefers our outputs 168:48 (p = 0.0195, subject-level sign test). The trained skill deploys as retrieval replay: 11.7 seconds per mesh on CPU, with no LLM calls.
查看此版本解读 →查看此版本原文 ↗