Developed a novel deep architecture for predicting future locations of people observed in first-person videos. Released a new dataset on human locomotion seen in FPV. Paper under review at CVPR’18.
Designed a knowledge distillation procedure aimed for compressing Fully Convolutional Network (FCN) with skip connections such as U-net and Stacked Hourglass Network. Achieved ~100x model compression.
Implemented various classfiers for human emotion detection in the Emotion Recognition in the Wild Challenge 2016 dataset to compare against results from CNN based approaches. Details on project page.