Search the book

Academic and professional experience, education, and selected research output.

Recent publications, presentations, projects, and opportunities.

Papers, datasets, software, and the research map.

IK-GAT predicts animation-ready joint orientations from 3D joint positions using a lightweight graph-attention network.

SIMSPINE uses biomechanics-aware simulation to provide vertebra-level 3D annotations and reference benchmarks for spine motion from RGB.

PoseAdapt provides benchmarks and a toolkit for adapting 2D human pose estimators across domain shifts and evolving skeletons.

SpinePose introduces SpineTrack and a biomechanics-aware approach to 2D spine pose estimation in unconstrained images and videos.

ClassroomKD adapts mentor selection and teaching intensity for each sample in image classification knowledge distillation.

SituationalLLM combines scene-graph representations and dialogue to support context-aware guidance for tasks in physical environments.

Shape2.5D combines synthetic and real textureless-object images with depth and surface-normal labels for estimation research.

Research on the SSBI bank-check signature dataset and a transformer-based method for detecting and verifying signatures.

CICA uses content-injected contrastive alignment to adapt CLIP for zero-shot and generalized zero-shot document image classification.

A unified skeleton and multi-teacher distillation approach to human pose estimation across datasets with different joint definitions.

Preliminary work on language-model-generated pose descriptions and FocusCLIP for zero-shot human-centric image classification.

Master's thesis on learning 3D reconstruction from images of textureless and transparent surfaces at TU Kaiserslautern.

A review of deep-learning methods for reconstructing 3D shape from a single RGB image, including textured and textureless surfaces.

BCSD provides 158 bank check images with pixel-level signature masks for training and evaluating signature segmentation.

Bachelor's thesis at NUST on automatically verifying handwritten offline signatures using image-based methods.

Apps, libraries, games, and software projects.

Open-source SpinePose CLI and Python API for spine-aware human pose estimation in images and video, with 2D and 3D model support.

An open-source Android port of OpenSim 4.6 with reusable Java and Kotlin APIs, GUI components, and a workbench app.

OpenDetect offers one Python interface for RF-DETR and YOLOX object detectors across ONNX Runtime backends, including person detection.

A look back at Hangman and Poppit!, two mobile games built around word guessing and timed balloon-popping play.

Earlier Android projects by Saif Khan: Doodle, RECatalyst, IPRD Geotag, Acacia Joinery, Scribe, and Tiled.

Interactive pose inference with WebAssembly and WebGPU.

Research collaboration topics and academic work.

Find a paper, dataset, tool, or collaboration route.

Completed theses, guided research, projects, and seminars.

A field guide to sensing, benchmarks, and open problems in spine research.

What spine vision tries to estimate, and why the answer depends on the task.

From radiographs and MRI to surface scans and ordinary video.

A comparison of public datasets, research releases, and clinical systems.

Gaps in validation, motion analysis, and practical inference.