9 new dissections, ontology v3.2, GitHub failure atlas
New
-
9 new paper dissections added across feed-forward 3D, classical SLAM, 3DGS, NeRF, and world models — closing gaps where the ontology referenced a paper but no dedicated page existed. Total dissections: 37 → 46.
- Feed-forward 3D origins: DUSt3R and MASt3R — the CroCo → DUSt3R → MASt3R → VGGT lineage now reads end to end.
- Feed-forward 3D frontier: π³ (Pi3) — permutation-equivariant, drops the reference-view assumption — and OVGGT — O(1) constant-memory streaming, currently the top production candidate in the streaming family.
- Classical SLAM: DSO (direct VO + photometric calibration) and SVO (semi-direct, up to 200 fps on embedded).
- 3DGS anti-aliasing: Mip-Splatting — drop-in fix for zoom-level aliasing artifacts.
- NeRF surfaces: NeuS — SDF + unbiased volume rendering bridge between radiance fields and meshes.
- World models: Aether — ICCV 2025 Outstanding Paper, first formal bridge between generative video (Sora / Cosmos / Genie) and 3D reconstruction (DUSt3R / VGGT).
-
GitHub failure atlas (§8.x) is now present on every new dissection. Each page surfaces real open issues, maintainer response times, and reproducibility caveats — so you know what breaks before you
git clone.
Improved
- Ontology upgraded to v3.2. The 5-axis taxonomy (problem / representation / sensor / paradigm / time) now covers all 46 dissections, and every page carries a machine-readable ontology header. Browse the taxonomy at
cheat-sheet/ontology. - Audit coverage expanded from 8/8 to 13/13 checks passing — atlas recall, 14-item template gate, 1,290+ internal link resolution, and ontology coverage all verified on every push.
- Page count refresh. Foundations now ships 92 deep dives across 199 pages. README badges, the foundations overview, and the feed-forward 3D overview reflect the new totals.
Notes
- Several new pages ship under non-commercial licenses upstream (DUSt3R / MASt3R / π³ weights are CC BY-NC; Mip-Splatting inherits INRIA 3DGS’s non-commercial clause). License flags are surfaced inline on each dissection — check before deploying.