Projects

Study language reconstruction from keywords and investigate the transferring of the models towards video captioning. Create a method of ranking sentences that better describe a video.

We studied recurrent models for multi-label prediction from frame features and the effect of global context. We investigated the usage of the labels correlation matrix before the final network prediction, in order to select the strongest cluster of labels.

Developed a method for detecting occluded zones in a video using a fully convolutional model with three branches for multi-model input, with feature alignment by optical flow. The resulting segmentation maps are refined by a second, more localised convolutional network.

An example of using the in-built project page.