People detection

People detection is a very challenging topic, where many top-level research groups are competing and have already proposed impressive approaches (e.g. Faster-RCNN, SSD, YOLO). Yet, we were able to design a novel algorithm able to better balance the speed and accuracy trade-off on the most challenging pedestrian detection benchmarks (e.g. Caltech and Citypersons).

Person Re-Identification

Person Re-Identification is a very challenging task, where current Computer Vision algorithms manage to obtain better results than humans. By proposing a simple and elegant technique, based on Spatial-Channel Partitions, we have obtained the best performance compared to the State-of-the-art approaches on the most popular benchmark datasets (e.g. Market-1501, CUHK03 and MARS).

Action recognition

This year, we have proposed several action recognition approaches able to outperform the State-of-the-art algorithms and get nearly maximal performance on most of ADL benchmark video datasets (e.g. Northwestern-UCLA Multiview Action 3D, NTUTU-RGB and DAHLIA). We have also released a novel ADL benchmark video dataset, which is more challenging, as it has been collected within real-world settings.

Awards