← mep

How we compare to the NBV literature

a comparative survey · 32 NBV papers · tagged by what is actually comparable to our interface

There is no shared public benchmark for next-best-view over BIM/MEP scenes: the NBV field spans object-centric turntables, RL drone policies, and NeRF-uncertainty planners, each with its own scene model, action space, and metric. So a raw number-vs-number ranking would be apples-to-oranges. Below, every published method is tagged by how well it maps onto our exact interface: a partial point cloud plus a candidate pose set, returning a per-pose score. comparable means same interface, reusable head or runnable baseline; partial means same spirit, mismatched scene/action model; off-pipeline means different inputs (RGB/NeRF) or output (RL actions, not candidate scores).

Where we sit. We do not claim SOTA on a public NBV benchmark: none exists for this setting. What we contribute is (a) a clean BIM/MEP eval harness over 297 manifest scenes plus IFC-Bench v2, (b) the OctoMap-IG baseline integration (Bircher et al., ICRA 2016) that our hybrid beats by +5.4σ / +5.0σ / +5.6σ across test_locked / held_out / OOD on surface coverage, (c) a metric-pathology fix in our own eval harness (with a 32-paper survey confirming the standard NBV convention already avoids it), and (d) the corpus > target finding: three target-engineering tricks regressed, one data-diversity fix won.
An evaluation trap we caught in our own pipeline. Our compute_mep_recall originally divided captured instances by the instances visible in the partial cloud: using the scanned cloud as its own ground truth. Observing one instance scored 1.0; oracle inflated 5-50×. We caught it because a newly-added extractor scored below random (random was inflated, the extractor was fine), and re-anchoring to scene-full GT (scene.instance_class) dropped oracle on gni_model_173 from 1.0 to 0.0093. We then surveyed the NBV literature to check whether this is a shared pitfall. The field standard is to normalize recall/coverage against the complete object or scene ground truth: object-reconstruction NBV (PC-NBV, NBV-Net, SCVP, GenNBV), info-gain NBV, semantic/instance NBV, and as-built/TLS scan planning all already do this. Across 32 papers we found no prior work with this pathology; the closest analog is the explicitly disclosed no-GT fallback in measurement-direct NBV (Surface Edge Explorer). So this was a real trap in our code, but we make no claim that other people's published numbers are affected (no contamination claimed).

The NBV field, tagged by comparability

Tagged against our interface: a (partial cloud, candidate poses) input that returns a per-pose score.

The one we reuse, and the one we beat. Our joint candidate-scoring head is PC-NBV-style (Zeng et al., IROS 2020): partial cloud in, per-candidate value out, so that line is genuinely comparable: same interface, we cite and reuse the design. The only method we run head-to-head on identical scenes is OctoMap-IG / nbvplanner (Bircher et al., ICRA 2016), integrated as our octomap_ig baseline; the learned-ranker-plus-lookahead hybrid beats it on every split. Everything else is the right idea on the wrong scene model (NBV-Net's 32³ grid plus fixed 14-view sphere; MA-SCVP's object-centric turntable) or a different problem entirely (NeU-NBV needs a per-scene NeRF; GenNBV samples RL actions, not candidate scores).

Corpus sources surveyed for expansion (Wave 6)

The OOD win came from more diverse buildings, not a cleverer loss, so Wave 6 surveyed open BIM/IFC corpora. IFC-Bench v2 was used (93 sub-scenes added); the rest are documented leads for further expansion.

IFC-Bench v2 · sylvainHellin/ifc-bench · CC-BY · USED (11 MEP IFCs, 93 sub-scenes) IfcOpenShell/files OpenIFC Auckland BIMData R&D buildingSMART Sample-Test-Files