CRF-Based Model for Instrument Detection and Pose Estimation in Retinal Microsurgery

Detection of instrument tip in retinal microsurgery videos is extremely challenging due to rapid motion, illumination changes, the cluttered background, and the deformable shape of the instrument. For the same reason, frequent failures in tracking add the overhead of reinitialization of the tracking. In this work, a new method is proposed to localize not only the instrument center point but also its tips and orientation without the need of manual reinitialization. Our approach models the instrument as a Conditional Random Field (CRF) where each part of the instrument is detected separately. The relations between these parts are modeled to capture the translation, rotation, and the scale changes of the instrument. The tracking is done via separate detection of instrument parts and evaluation of confidence via the modeled dependence functions. In case of low confidence feedback an automatic recovery process is performed. The algorithm is evaluated on in vivo ophthalmic surgery datasets and its performance is comparable to the state-of-the-art methods with the advantage that no manual reinitialization is needed.