Sentinel

Sentinel is a Python framework for evaluating occlusion-centered next-best-view (NBV) methods. It provides immutable trial and method contracts, candidate generation and scoring, source-only geometry state, evaluation oracles, and experiment utilities.

Install

Requires Python 3.11 or newer.

python -m venv .venv
source .venv/bin/activate          # POSIX
# .\.venv\Scripts\Activate.ps1    # PowerShell
python -m pip install -r requirements-public.lock
python -m pip install -e .

Install optional dependencies only when needed:

python -m pip install -e ".[viewer]"  # local geometry viewers
python -m pip install -e ".[render]"  # rendered image loading
python -m pip install -e ".[oracle]"  # mesh coverage utilities

The RGB-to-geometry adapters load their model runtimes lazily. Install the corresponding upstream runtime and model package before using MoGe-2 or Depth Anything 3.

requirements-public.lock records the verified base runtime used for the appendix example. Optional viewer, rendering, oracle, and model-runtime dependencies are intentionally separate.

Quickstart

Run the self-contained example:

python quickstart.py

quickstart.py generates an occlusion-centered candidate set, builds a trial, projects it to the method-visible MethodInput, ranks the candidates, validates the MethodDecision, and creates an evaluation-only EvaluationBundle.

Package areas

• sentinel/trial: immutable scenes, views, trials, method inputs, decisions, and evaluation bundles.
• sentinel/nbv: candidate generators, baselines, and NBV scoring.
• sentinel/state: source-only pseudo-3D state and voxel products.
• sentinel/oracle and sentinel/metrics: evaluation utilities and aggregate metrics.
• sentinel/render, sentinel/pipeline, and sentinel/experiment: render-cache, pipeline, and experiment-run support.
• sentinel/backbone and sentinel/visualization: geometry adapters and local viewers.

Command-line entry points

sentinel-run --help
sentinel-reconstruction-viewer --help
sentinel-failure-annotations --help

The public bundle does not include datasets, render scripts, model weights, or generated experiment artifacts. The full pipeline entry point therefore requires those external resources to be supplied separately; the quickstart is the self-contained public smoke test.

Citation

See CITATION.cff for citation metadata.

License

See LICENSE. This is source-available proprietary software; no redistribution rights are granted by this repository.