NBV-Bench (ray-coverage MVP)

A small, method-agnostic Next-Best-View (NBV) benchmarking starter with realistic ray-visibility coverage (Open3D), area-uniform hemisphere candidates, and a lean run loop that logs just what you need to compare methods.

Goal: a clean, reproducible baseline where adding new selectors (Greedy, Random, PSO, SMA, …) is trivial and results are directly comparable.

---

Project layout

assets/
  tsdf_mesh.ply            # test mesh/asset
experiments/
  configs/baseline.yaml    # balanced defaults
  results/                 # metrics + selected-indices CSVs

nbvbench/
  core.py                  # main loop (minimal outputs)
  data.py                  # mesh loading + hemisphere candidates (area-uniform)
  geo.py                   # RayCoverage (Open3D), AUC helpers
  selectors.py             # Greedy, Random; drop-in for new methods
  scorers.py               # interfaces to geo coverage

visualize_frustums_o3d.py  # white-bg viz: aim/frusta/both (spheres)
CHANGELOG.md               # versioned changes
README.md                  # this file

What’s in this MVP (v0.2)

• RayCoverage (Open3D tensor raycasting) with real occlusion: frustum gating (fx/fy/W/H), near/far planes, optional front-facing filter, and eps hit tolerance over GT surface samples.
• Area-uniform hemisphere sampling over a φ-band (z-stratified spherical Fibonacci) to avoid “top-ring” bias.
• Lean outputs only:
  • experiments/results/metrics_<tag>_<selector>.csv (per-step coverage)
  • experiments/results/<tag>_<selector>_selected.csv (chosen indices)
• Visualizer (visualize_frustums_o3d.py): white background; gray candidates; green selected spheres; aim lines and/or frusta; sphere sizes ∝ scene extent; unit scaling; auto-fit.
• Balanced baseline.yaml with realistic defaults.

See CHANGELOG.md for v0.1 → v0.2 details.

---

Requirements

• Python 3.9+ (tested)
• open3d, numpy, pyyaml
• A triangle mesh (e.g., assets/tsdf_mesh.ply)

Install:

pip install -r requirements.txt

Quickstart

Run Greedy or Random on the baseline config:

# Greedy
python -m nbvbench.core --cfg experiments/configs/baseline.yaml --selector greedy

# Random
python -m nbvbench.core --cfg experiments/configs/baseline.yaml --selector random

Visualization

Visualize a run (white background, gray candidates, green selected; aim lines + frusta):

python visualize_frustums_o3d.py \
  --cfg experiments/configs/baseline.yaml \
  --indices_csv experiments/results/baseline_greedy_selected.csv \
  --mode both --recenter --axes --auto_fit \
  --cand_glyph spheres \
  --cand_radius_factor 0.015 \
  --sel_radius_factor 0.03 \
  --frustum_scale_factor 0.8

Roadmap

• TSDF coverage (raycast depth → integrate → nearest-distance coverage)
• Additional selectors: PSO, SMA baselines
• Multi-object scenes / presets