In collaboration with Prof J Benois-Pineau (LABRI), we develop a line of research aiming at integrating tools from computer vision augmented with gaze information, in order to help prosthesis and robotics controls. In particular, we improved automatic object detection using a Deep Convolutional Neural Network (DeepCNN, Fig3A) operating on egocentric video at eye level in conjunction with saliency map from gaze information (Fig3B-D) (Pérez de San Roman et al., 2017). By combining CNN and LSTM (Long Short Term Memory) architecture, we were also able not only to detect object localization, but also to predict the intention to grasp an object, and this in natural environments as complex as real kitchens (Fig3E) (Gonzalez-Diaz et al. 2019).
This information from computer vision and gaze information is critical to the alternative kinematic-based control developed in virtual reality Theme 5, and provide rich perspectives of integration on our robotic platform REACHY Theme3, which can now host computer vision algorithms.
Publications: Gonzalez-Diaz I, Benois-Pineau J, Domenger J-P, Cattaert D, de Rugy A (2019) Perceptually-guided deep neural networks for ego-action prediction: Object grasping. Pattern Recognition. 88: 223-235.
Pérez De San Roman P, Benois-Pineau J, Domenger J-P, Cattaert D, Paclet F, de Rugy A (2017) Saliency Driven Object Recognition in Egocentric Videos with Deep CNN: toward application in assistance to Neuroprostheses. Computer Vision and Image Understanding. 164, 82-91.