Soroush Mehraban

I received my PhD from the University of Toronto, where I was advised by Dr. Babak Taati and was a Faculty Affiliate Researcher at the Vector Institute. I am currently a Member of Technical Staff (Research) at Bagel Labs, where I work on Physical AI. Previously, I interned at Pickford AI, where I worked on video style transfer, 3D gesture generation, and text-to-3D scene generation. My research spans generative models, computer vision, and human motion analysis, including 3D human pose estimation, 3D human mesh recovery, action recognition, and gait assessment.

Email  /  CV  /  Scholar  /  Linkedin  /  YouTube  /  Github

Research

My research focuses on generative models and their applications across a range of computer vision tasks. Below, I highlight some of my recent contributions.

SynthGait-19K is a physically grounded synthetic dataset of 19,272 walking videos derived from 6,427 motion-capture sequences across 437 subjects. It pairs diverse RGB videos with SMPL motion and six clinically meaningful gait parameters, and introduces GaitXFormer for direct gait-parameter estimation from video.

FastHMR introduces two merging strategies, Error Constrained Layer Merging (ECLM) and Mask guided Token Merging (Mask ToMe), to reduce computational cost and redundancy in transformer based 3D Human Mesh Recovery. ECLM selectively merges layers with minimal impact on MPJPE, while Mask ToMe merges background tokens that contribute little to prediction. A diffusion based decoder further enhances performance by using temporal context and pose priors. The method achieves up to 2.3x faster inference while slightly improving accuracy across benchmarks.

Puppeteer generates posture-aware, object-grounded co-speech gestures in a causal latent space.

PickStyle is a diffusion-based video style transfer framework that preserves video context while applying a target visual style. It uses low-rank style adapters and synthetic clip augmentation from paired images for training, and introduces Context-Style Classifier-Free Guidance (CS-CFG) to independently control content and style, achieving temporally consistent and style-faithful video results.

Pain in 3D introduces 3DPain, a large-scale dataset of controllable synthetic pain faces with paired neutral references, action-unit annotations, and clinically grounded pain scores. It also presents ViTPain, a reference-guided vision transformer for identity-aware automated pain assessment.

A simple yet effective method based on token shuffling for extending the benefits of CFG to broader settings, including unconditional generation.

LIFT enables unified implicit neural representations across diverse tasks by leveraging localized implicit functions and a hierarchical latent generator.

GAITGen is a generative framework that synthesizes realistic gait sequences conditioned on Parkinson’s severity. Using a Conditional Residual VQ-VAE and tailored Transformers, it disentangles motion and pathology features to produce clinically meaningful gait data. GAITGen enhances dataset diversity and improves performance in parkinsonian gait analysis tasks.

STARS enhances the Mask Autoencoder (MAE) approach in self-supervised learning by applying contrastive tuning. We also show that MAE approaches fail in few-shot settings and achieve improved performance by using the proposed method.

Evaluating recent motion encoders for the task of parkinsonism severity estimation (UPDRS III gait)

Estimating 3D locations of 17 main joints from a monocular video.

Website Template