"Was shocked to find our original engine would take centuries to generate the dataset we needed, but we pulled through... An awesome virgin experience to multimodal AI research!"
”
#Generative AI Researcher Intern
SynDORBench
Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions
[+] Engineered a scalable Python and Blender synthetic-data engine with configurable scenes, environmental conditions, camera settings, and automated pose, bounding-box, and segmentation-mask labels using SMPL-X, AMASS, and BABEL.
[+] Built a unified evaluation framework with Hydra configs and Docker/CUDA for human-presence classification and action recognition. Compared YOLO with compact LVLMs on human presence; action recognition evaluated LVLMs only.
