Visual Tracking of Surgical Tools for Proximity Detection in Retinal Surgery

Data-Driven Visual Tracking in Retinal Microsurgery

“…From the figures, we clearly see that our approach outperforms that all three trackers. For example, for δ = 20 our approach detects over 70% more than [6] and over 40% more than [10]. This corresponds to 449 and 309 more correct instrument detections, respectively.…”

“…To work reliably, gradient-based trackers [6,9] need continuous template updating to maintain accurate position estimation when changes in the target appearance are severe. 2.…”

“…We compare our approach to three existing gradient based trackers on the same set of images: the Mutual Information of [6], the SCV of [10] and the SSD tracker used in this pipeline. To allow a fair comparison, when any of these trackers provided false detections they were re-initialized with the ground truth, and we report the proportion of frames where re-initialization is not required.…”

“…Most existing methods have relied instead on knowing the instrument geometry beforehand to solve complex optimization problems [3,4], or have constructed sophisticated and robust objective functions within more traditional gradient-based frameworks to deal with appearance change [5,6]. Typically, these methods have extremely simple appearance models that combine geometry with colour or edge-based features, and ultimately work well only in limited conditions such as in eye phantoms.…”

“…Typically, these methods have extremely simple appearance models that combine geometry with colour or edge-based features, and ultimately work well only in limited conditions such as in eye phantoms. For example, using the method of [6], tracking is often lost after only 5 frames in the in-vivo sequence of Fig. 1.…”