From Bob Wiki, a public encyclopedia-style profile page
This page is not a ranked list of Bob’s favorite papers. It records the technical vocabulary that shaped how Bob thinks about 3D/4D reconstruction, neural representations, animals, and embodied AI.
Editorial note. Bob does not describe his research taste as coming from a single “life-changing” paper. It emerged from repeatedly studying many papers, codebases, demos, and technical trends around implicit representation, neural rendering, generative 3D, animal priors, and dynamic reconstruction.
Geometry can be represented as a learned continuous field rather than only as explicit meshes, voxels, or point clouds.
This was Bob’s entry point into 3D reconstruction. Before DogRecon, he was interested in implicit neural representations and monocular 3D reasoning.
These works made 3D reconstruction feel like a language for converting images into possible 3D worlds. Bob’s later animal work still carries this intuition: a single observation is not enough by itself, but it constrains a structured space of plausible 3D bodies.
View synthesis can become a reconstruction engine. Rendering, optimization, appearance, and geometry can be learned together.
Helped connect Bob’s interest in 3D geometry with image formation and visual realism. This line of work shaped the background for thinking about renderable animal and human avatars.
3D Gaussian representations
Representative papers / methods
What Bob learned
Connection to Bob’s work
3D Gaussian Splatting; dynamic Gaussian methods; Gaussian avatar methods
A 3D representation can be explicit, efficient, renderable, and suitable for animation or editing.
Directly connected to DogRecon and AniGauss. DogRecon uses animatable 3D Gaussian dog models, and AniGauss uses Gaussian representations as an anchor representation during inference.
For Bob, the importance of Gaussians is not only speed or rendering quality. They also suggest a practical representation layer for controllable bodies: something that can be optimized, rendered, animated, and connected to downstream systems.
Generative models can provide missing views, priors, or 3D hypotheses from a single image.
DogRecon was shaped by this general direction, but it also exposed the reliability problem: generated views are not always consistent or correct. This motivated canine-prior and mask-guided constraints, as well as careful sampling among generated views.
This category is especially important because it shows Bob’s relationship to generative AI: useful, powerful, but not automatically reliable. A reconstruction system needs priors, constraints, and selection mechanisms to convert generated evidence into stable 3D structure.
Human reconstruction is a mature testbed for pose, deformation, identity preservation, appearance modeling, and controllable animation.
Bob asks which assumptions transfer from humans to animals and which fail. This comparison helps position animal reconstruction as a harder and less standardized extension of animatable body modeling.
Animal body modeling
Representative papers / methods
What Bob learned
Connection to Bob’s work
SMAL; BARC; WLDO-style animal pose/reconstruction; LASR; BANMo; MagicPony; animal avatar and animal pose literature
Animals need category-specific priors, weak supervision, flexible shape modeling, and methods that tolerate limited datasets and ambiguous observations.
This is the direct path to DogRecon, AniGauss, and WildAni4D. Bob’s work can be read as an attempt to make animal reconstruction more controllable, more general, and eventually more dynamic.
The animal literature matters because it makes the limitations of human-centric reconstruction visible. Dogs, horses, cats, and other animals are not simply humans with different skeletons. They require different priors, different data assumptions, and different evaluation habits.
Dynamic 4D reconstruction
Representative papers / methods
What Bob learned
Connection to Bob’s work
D-NeRF; Nerfies; HyperNeRF; Neural Scene Flow Fields; dynamic 3D Gaussians; 4D Gaussian Splatting; video-based human and object reconstruction methods
Reconstruction should include time, deformation, motion, and trajectory, not only static shape.
This supports Bob’s shift from static object/scene reconstruction toward dynamic animal reconstruction. WildAni4D is the clearest project in this direction.
Physics-grounded motion and embodied characters
Representative papers / methods
What Bob learned
Connection to Bob’s work
PhysHMR; PhysCtrl; ModSkill; Vid2Sim; physics-based human motion reconstruction and control; simulation-to-reconstruction methods
Visual plausibility is not enough. Bodies should move in ways that respect contact, balance, dynamics, and physical constraints.
This is a future-facing direction for Bob: physics-grounded 4D animal reconstruction from in-the-wild videos. It connects his animal reconstruction background with motion, simulation, and embodied AI.
This category is especially relevant to Bob’s current visiting-research framing. The open question is how ideas from physics-grounded human motion can be translated to non-human articulated bodies, where morphology, gait, data availability, and priors differ substantially.
Open bibliography to fill later
This page should eventually become a more concrete bibliography. Each entry can follow a simple format:
Paper / method
Why it mattered
Connection to Bob
Paper title
One-sentence technical influence.
Which project or research question it shaped.
Candidate entries to verify
NeuS and VolSDF for neural surface reconstruction.
Convolutional Occupancy Networks and DeepSDF for implicit shape representation.
NeRF, Mip-NeRF, and Instant-NGP for neural rendering.
3D Gaussian Splatting and dynamic Gaussian methods for efficient renderable representations.
Zero123 and related single-image 3D generation methods for missing-view priors.
SMAL, BARC, BANMo, LASR, and MagicPony for animal shape and reconstruction.
D-NeRF, Nerfies, HyperNeRF, and 4D Gaussian methods for dynamic reconstruction.
PhysHMR, PhysCtrl, ModSkill, and Vid2Sim for physics-grounded reconstruction and motion.
Next step: Bob can mark each candidate as “strong influence,” “background knowledge,” or “remove.”