University of British Columbia
A unified framework for disentangled 3D mouse body-state estimation
Abstract
dc:descriptionAnimal behaviour is a central readout in neuroscience and disease-model research. Video provides a practical way to record naturalistic and task-driven behaviour in laboratory mice. However, the scientific meaning of video-derived measurements depends on the representation used to describe the animal. Modern markerless methods can recover 2D/3D landmarks, silhouettes, 3D surfaces, or learned behavioural embeddings. These representations are useful but incomplete descriptions of body state: sparse landmarks omit surface shape and articulation, silhouettes and 3D surfaces do not identify anatomical structure, and learned behavioural embeddings may entangle behaviour with viewpoint, appearance, and recording conditions. This thesis develops an anatomically grounded body-state representation for video-based mouse behaviour analysis. The main contribution is PokeMouse (Parametric optimizable kinematic embedding for Mouse body-state inference), a mouse-specific framework that combines an articulated parametric body model with an optimization-based inference procedure. PokeMouse represents the animal through explicit variables for global position, root orientation, local joint rotation, body-part scale, subject morphology, and residual surface deformation. Rather than treating keypoints, masks, or surfaces as final behavioural measurements, it treats them as observations that provide evidence about an underlying latent body state, constrained by anatomical priors. Here, "anatomically grounded" refers to model construction using a CT-informed template and anatomical priors. Validation covers surface geometry, sparse surface landmarks, and image-space consistency; recovered joint centres and rotations have not been independently validated. PokeMouse was evaluated using CT-derived surfaces, a public RGB mouse benchmark, and laboratory-collected infrared grayscale recordings. Fitting to CT-derived surfaces produced a mean symmetric Chamfer distance of approximately 0.8 mm. On the RGB benchmark, PokeMouse achieved lower sparse-landmark error than representative prior methods under a shared eight-keypoint protocol. Held-out camera evaluation showed that fitted states explained unseen image evidence with low reprojection error and high silhouette agreement, while infrared recordings demonstrated usability under imaging conditions closer to behavioural neuroscience. Because iterative inverse-graphics fitting is computationally expensive, this thesis introduces PokeDex (PokeMouse Discrete-pose Explicit-state estimator), a neural inference module that predicts articulated PokeMouse body states from single-view video using DINO-derived visual features. Together, PokeMouse and PokeDex provide an explicit, anatomically informed body-state representation with potential for downstream behavioural analysis.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Cell and Developmental Biology
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hu, Hao
- Contributors dc:contributor
-
- Murphy, Timothy H.
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95747