Visual object tracking has experienced a quantum performance leap in the last
decade with adoption of deep learning. While most of the work has been dedicated
to opaque object tracking, much less research has been invested in tracking of
transparent objects. One important reason is the lack of appropriate training
datasets. We addressed this by proposing a new dataset for training deep models
for transparent object tracking. We utilize rendering to acquire challenging
sequences with attributes not present in opaque object datasets. We show that
re-training state-of-the-art trackers on the new dataset leads to up to 16%
performance boosts on the recent transparent object tracking benchmark. We will
discuss several insights in the talk, such as why training exclusively on
transparent objects leads to inferior performance, what does a transformer
tracker rely on when tracking transparent objects, and how the performance is
affected by backbone depth.