CVPR 2016

First Person Action Recognition Using Deep Learned Descriptors [project page] [PdF] [code] [notes]

Suriya Singh, Chetan Arora, C. V. Jawahar

Synthesis

Contributions

• demonstrate the utility of deep learned egocentric features for the first person action recognition and show that these features are complementary to state of the art flow and image features

• train an ‘Ego ConvNet’ to learn the coordination patterns between hands, head and eye movement

• learn hand/light model from small set of manually segmented hand regions in diverse imaging conditions

• create appearance descriptors

• with various Gabor filters

• Using appearance descriptors