M2M-HMR
Monocular Models are Strong Learners for Multi-View Human Mesh Recovery
Applied Scientist II · Amazon Music
Muhammad Usama Saleem
Scholar GitHub LinkedIn X Email
I'm an Applied Scientist II at Amazon Music, where I work on multimodal content understanding for music, including track and artist understanding across all genres, as well as sonics, sequencing, and recommendation, building LLM-based systems with LoRA fine-tuning and automated evaluators. I received my Ph.D. in Computer Science from the University of North Carolina at Charlotte in 2026, advised by Prof. Pu Wang in the GENIUS Lab. I was previously a Research Scientist Intern at Google's Extended Reality (AR/VR) team, and my work includes a U.S. Patent on 3D human pose and movement estimation from monocular images.
Research interests
3D Computer Vision·Digital Humans·Generative Motion·Diffusion Models·Multimodal Learning·Audio & Music Representation·Interactive XR
Open to research collaborations and opportunities. Reach me at usama.saleem7977@gmail.com.
Experience
2026 - Present
Applied Scientist II · Amazon Music · Seattle, WA
2025 - 2026
Research Scientist Intern · Extended Reality (AR/VR) · San Francisco, CA
Summer 2025
Research Scientist · Multimodal GenAI · Boston, MA
2023 - 2025
Research Lead · Charlotte, NC
News
-
Aug 2026
Real-Time Neural Musculoskeletal Pose Estimation demo paper published at IEEE/ACM CHASE 2026.
-
Jun 2026
MAGE on modality-agnostic music generation released on arXiv.
-
Jun 2026
M2M-HMR accepted to ECCV 2026 (Oral).
-
Mar 2026
Joined Amazon as Applied Scientist II (Full-Time).
-
Mar 2026
Completed Ph.D. from UNC Charlotte. Dissertation: Towards Generative Modeling of 3D Humans From Multimodal Signals.
-
Mar 2026
U.S. Patent published: 3D human pose and movement from monocular images.
-
Feb 2026
LiveGesture accepted to CVPR 2026.
-
Nov 2025
DanceMosaic accepted to AAAI 2026.
-
Oct 2025
Joined Google as Research Scientist Intern, Extended Reality (AR/VR).
-
Oct 2025
MaskControl selected for Oral Presentation & Best Paper Award Nominee at ICCV 2025.
-
July 2025
Poster selected at Amazon WWAS Science Fair Seattle.
-
June 2025
MaskHand & MaskControl accepted to ICCV 2025.
-
Dec 2024
GenHMR accepted to AAAI 2025, presented in Philadelphia.
-
Oct 2024
BioPose accepted to WACV 2025.
-
July 2024
BAMM accepted to ECCV 2024.
-
Sept 2023
Joined Lowe's as Research Lead of the CV & UNCC Team.
-
June 2023
DPFedProxGAN accepted to IJCNN 2023.
-
July 2022
Privacy Enhancement for Few-Shot Learning accepted to IJCNN 2022.
-
Jan 2022
DP-Shield accepted to EDBT 2022.
Selected Research
LiveGesture
Streamable Co-Speech Gesture Generation Model
Walk Before You Dance
High-fidelity and Editable Dance Synthesis via Generative Masked Motion Prior
MaskHand
Generative Masked Modeling for Robust Hand Mesh Reconstruction in the Wild
MaskControl
Spatio-Temporal Control for Masked Motion Synthesis
with
GenHMR
Generative Human Mesh Recovery
BioPose
Biomechanically-accurate 3D Pose Estimation from Monocular Videos
BAMM
Bidirectional Autoregressive Motion Model
Publications
2026
Methods, Systems, and Computer Program Products for Generating 3D Human Pose and Movement Estimation from Monocular Image Information
U.S. Patent, US20260065565A1, 2026
MAGE: Modality-Agnostic Music Generation and Target-Source Extraction
In collaboration with
arXiv, 2026
Work done during research internship at Google.
Walk Before You Dance: High-fidelity and Editable Dance Synthesis via Generative Masked Motion Prior
AAAI, 2026
*Equal Contribution
2025
MaskControl: Spatio-Temporal Control for Masked Motion Synthesis
In collaboration with
ICCV, 2025Oral · Best Paper Award Nominee
BioPose: Biomechanically-accurate 3D Pose Estimation from Monocular Videos
WACV, 2025
*Equal Contribution
2024
2023
2022
Research Focus
My research spans three areas: generative 3D human reconstruction, controllable motion synthesis, and multimodal audio generation and separation, with collaborations across Google DeepMind/XR, Amazon, Snapchat, and Lowe's.
Generative 3D Human Reconstruction
Generative models that recover accurate 3D human pose and mesh directly from monocular images and video, replacing single-prediction regressors with masked and diffusion-based generation to handle depth ambiguity and occlusion. This spans full-body mesh recovery (GenHMR, M2M-HMR), hand mesh reconstruction in the wild (MaskHand), and biomechanically accurate pose estimation grounded in anatomical constraints (BioPose), alongside a U.S. patent on 3D pose and movement estimation from monocular images.
Controllable Motion Synthesis
Generative frameworks for controllable, high-fidelity 3D human motion: streamable co-speech gesture generation with zero look-ahead (LiveGesture), editable dance synthesis via a generative masked motion prior (Walk Before You Dance), spatio-temporal control over masked motion synthesis (MaskControl, with Snapchat), and bidirectional autoregressive motion modeling (BAMM). I'm interested in how masking, tokenization, and diffusion jointly enable real-time, controllable generation.
Multimodal Audio Generation and Separation
Extending generative modeling beyond vision to audio and cross-modal settings, including a modality-agnostic framework for conditional music generation and mixture-grounded target-source extraction (MAGE), developed during a research internship at Google. At Amazon Music, I work on multimodal content understanding for music, including track and artist understanding across all genres, sonics, sequencing, and recommendation, using LLMs with LoRA fine-tuning and automated evaluators. I'm broadly interested in how the same generative principles (masking, tokenization, diffusion) transfer across modalities to build unified multimodal systems.