Research
I build methods that turn images and videos into animatable human and animal representations. [publications]
Research trajectory
- 2022–2023: depth and fisheye geometry → SlaBins (ICCV 2023), with 42dot. [paper]
- 2024: single-image animatable dogs → DogRecon, CV4Animals oral. [work]
- 2025: DogRecon (IJCV), Avatar++, Samsung display-inspection collaboration (2024–2025). [papers] [industry]
- 2026: WildAni4D, AniGauss, UT Austin Research Intern, Sony collaboration. [papers] [collaboration]
DogRecon
What I built: Single-image dog reconstruction with canine-prior-guided novel views and an animatable Gaussian representation. [project]
What it showed: Internet-sourced image-to-3D test: CLIP-Similarity 0.8006. This is a rendered-view semantic-similarity metric, not geometry accuracy. [Table 2]
Status: IJCV 2025; CV4Animals @ CVPR 2024 oral. [paper record]
Links: [figure / demo]
WildAni4D
What I built: Video-native recovery of animal mesh, articulated motion and world-space trajectory. [project]
What it showed: Synthetic test set: MPJPE 91.29 and Shape Consistency 0.0703 (lower is better). These are paper-reported values; units are not specified in the table. [Table 1]
Status: CVPR 2026 Findings. [paper record]
Links: [figure / demo]
AniGauss
What I built: Topology-aware Gaussians for animatable animal reconstruction from a single in-the-wild image. [project]
What it showed: The published workshop manuscript reports qualitative comparisons across animal categories; no quantitative performance claim is made here. [Figure 2]
Status: CV4Animals Workshop at CVPR 2026. [paper record]
Links: [figure / demo]
Fisheye depth — SlaBins
What I built: contributed to slanted-bin fisheye depth estimation for road environments. [paper]
What it showed: SynWoodScape test results: RMSE 1.040 and Abs Rel 0.022. These results refer to this benchmark, not production validation. [Table 1]
Status: ICCV 2023. [paper / project / augmentation code]
Manuscripts & Ongoing Research
Working project titles; listed separately from published work.
- AniGauss++ — Manuscript submitted to IJCV.
- Multi-Dancer Generation — Manuscript in preparation for CVPR. Research collaboration with Sony × UT Austin.
- Animal Agent — Manuscript in preparation for CVPR.
- Animal Motion — Manuscript in preparation for CVPR.
Origin story — 2022–2024
Bob first became interested in 3D reconstruction through implicit neural representations and monocular depth estimation. Early influences included neural surface reconstruction and occupancy-based representations, such as NeuS-style neural surface modeling and Convolutional Occupancy Networks. These works made 3D reconstruction feel like a way to connect geometry, learning, and visual understanding into a single technical language.
Over time, Bob became less satisfied with only reconstructing static objects or scenes. Static reconstruction felt important, but he expected many of the core problems to be solved relatively quickly. He became more attracted to dynamic reconstruction: understanding bodies that move, deform, interact with the ground, and change over time.
The shift toward animals came from a mixture of technical opportunity and long-term motivation. Animal reconstruction was much less explored than human reconstruction, yet animals are everywhere in the real world. If the goal is to build digital twins or embodied AI systems that reflect the real world, animals cannot be ignored. They are also technically harder: they have diverse body shapes, less standardized data, weaker priors, fur and appearance ambiguity, and complex locomotion. This made animal 3D/4D reconstruction feel like both an underexplored research space and a more challenging testbed for general dynamic reconstruction.
Before DogRecon — 2022–2024
Before DogRecon, Bob was mainly interested in implicit neural representations and monocular depth estimation. This background shaped how he thought about geometry from limited visual evidence: a single image should not be treated as a flat observation, but as a constraint on possible 3D worlds.