In the context of retinal microsurgery, visual tracking of instruments is a key component of robotics assistance. The difficulty of the task and major reason why most existing strategies fail on {\it in-vivo} image sequences lies in the fact that complex and severe changes in instrument appearance are challenging to model. This paper introduces a novel approach, that is both data-driven and complementary to existing tracking techniques. In particular, we show how to learn and integrate an accurate detector with a simple gradient-based tracker within a robust pipeline which runs at framerate. In addition, we present a fully annotated dataset of retinal instruments in {\it in-vivo} surgeries, which we use to quantitatively validate our approach. We also demonstrate an application of our method in a laparoscopy image sequence.