RESEARCH_BACKLOG.md

Simam3D research backlog

This backlog separates testable hypotheses from implemented behavior. None of the ideas below should be described as a result until an experiment produces a held-out metric and a provenance-validated artifact.

1. Evidence-guided hidden-side completion

Hypothesis: conditioning a completion model on visible depth, confidence, lighting direction, and the user's scene hypothesis reduces implausible back-side geometry compared with image-only completion.

Experiment: hold out one or more views from a calibrated object capture; compare image-only, depth-only, and depth-plus-evidence conditions. Report held-out depth error, silhouette IoU, novel-view perceptual error, and failure rate on occluded markings. The user's hypothesis must be logged as metadata, never treated as ground truth.

2. Uncertainty-guided view acquisition

Hypothesis: selecting the next camera from the largest uncertainty-weighted surface gap improves coverage per captured view over a fixed azimuth schedule.

Experiment: run both policies on the same synthetic and real multi-view sets. Measure coverage, reprojection consistency, held-out depth, and points with contributions from at least two source views after each added frame.

3. Photometric light-field consistency

Hypothesis: a depth-normal light estimate can reject inconsistent Gaussian colors before optimization, improving relit-view stability without changing geometry.

Experiment: compare no filtering, image-space light filtering, and depth-normal filtering under controlled lighting. Report color error under held-out relighting and geometry preservation. The current detector is only a heuristic diagnostic and is not a lighting reconstruction result.

4. Cross-modal reference-to-avatar handoff

Hypothesis: a capture manifest containing identity-safe labels, view hashes, coverage, and explicit consent improves reproducibility of reference-conditioned video generation compared with an unstructured prompt and image folder.

Experiment: compare reference sets using the same approved subject and prompt. Track reference retention, temporal identity drift, pose adherence, manual artifact review, and disclosure/watermark compliance. Do not publish identifiable samples without permission.

5. Gaussian evidence instead of opaque splats

Hypothesis: retaining confidence, source-view pointers, and contribution bitmasks in the Gaussian initialization makes downstream pruning and debugging more reliable than an unlabeled point cloud.

Experiment: ablate the evidence fields and measure removal of low-confidence points, source disagreement, memory, render quality, and held-out-view error. The current implementation provides the evidence fields and deterministic fusion; it does not yet perform learned Gaussian optimization.

6. Calibration-aware fusion gating

Hypothesis: combining DA3 camera poses with an explicit view-diversity and cross-view-support gate reduces false confidence from synthetic orbit priors, especially when users provide repeated, near-duplicate, or weakly overlapping views.

Experiment: evaluate three fusion modes on calibrated captures: synthetic orbit priors, native DA3 poses, and native DA3 poses with low-support pruning. Hold out one view and report reprojection coverage, relative-depth error, silhouette overlap, and the fraction of retained points supported by two or more distinct inputs. Include duplicate and near-duplicate controls so a high overlap score cannot be mistaken for useful view coverage.

7. Observed execution evidence (not a quality result)

The hosted v0.4.1 pipeline has completed one-, two-, and three-input synthetic requests with native DA3 provenance and direct Gaussian export. The controlled three-input stress test recorded 9.9% multi-view voxel support, while a duplicate-input request was explicitly labeled duplicate_input_support. These observations motivate the calibration experiment above; they are not evidence of metric accuracy, unseen-side correctness, or superiority over a baseline.

Current acceptance rule

Every completed experiment must include: pinned code and model revisions, input hashes, runtime/hardware, dataset split, baseline, metrics, failure cases, and a manifest validated by validate_manifest.py. A visually convincing turntable alone is insufficient evidence for hidden-surface accuracy or a "perfect digital twin."