Seol mar théacs é seo: Cross-modal learning from visual information for activity recognition on inertial sensors