Abolfazl (Alex) Meyarian

Ongoing

• Direct Preference Optimization for Text-to-Video/Text-to-Image generation.

Published

SPADe: Spatial Plaid Attention Decoder for Semantic Segmentation

• Created an efficient and powerful decoder for semantic segmentation called “SPADe”, compatible with CNNs and Transformers as the backbone.

• Achieved state‑of‑the‑art performance on Cityscapes and ADE20K with a significantly reduced computational cost of 34.95% and 32.85% lower than the second‑best decoder, for each dataset, respectively.

• Minimized decoder size by 89.74% relative to UPerNet.

Irrigation Practice Mapping

• Developed advanced AI models to classify irrigation practices in U.S. lands accurately. 

• Managed data acquisition, annotation, and preprocessing, as well as designing, training, and deploying the models across a large image dataset.

• The proposed model could segment irrigation practices with a mIoU of over 93.0%.

• Provided a nationwide irrigation map to USGS and NASA, providing insight into water use across the U.S.

A Task-Aware Network for Multi-task Learning 

• Designed a task‑aware multi‑task network for landmark localization, pose estimation, and landmark visibility estimation.

• Reduced error by 25% for face pose estimation, 15% for landmark localization, and 10% for landmark visibility estimation with a novel multi‑task network.

• Secured the best cumulative performance with a 5.7% error reduction, despite landmark visibility being the most challenging task.

• Ensured superior consistency with smaller standard deviations in error rates across tasks.