Planar object tracking benchmark in the wild

Visual Coin-Tracking: Tracking of Planar Double-Sided Objects

“…In this section we show that the proposed CTR-Base method outperforms general state-of-the-art trackers on the CTR dataset and retains good performance on the POT-210 [18] dataset. Then we demonstrate that the homography-based pose modeling prevents the CTR-Base tracker from making fatal mistakes.…”

“…The mean IoU (mIoU) is 0.81, showing that our method generalizes to POT-210 well. The best results were achieved on the outof-view and the perspective distortion subsets of [18] with mIoU 0.89 and 0.88 respectively, while the worst on the motion blur subset with mIoU of 0.71.…”

“…Therefore, there is a significant domain gap between these datasets and the proposed coin-tracking problem. Other datasets for tracking planar object exist, such as [18,5], but they only contain sequences with single side of the planar object visible. Moreover, in most cases the objects are fixed and the camera moves around them.…”

“…The planar object tracking datasets [5,18] are the closest to the CTR dataset, but they only contain a single sided view of the object; the viewing angle range is limited. In most of the sequences the tracked object is fixed to the background behind it, e.g.…”

“…Computing IoU from the bounding boxes might not seem fair, but the performance gap demonstrates the need of representing the tracked object by segmentation, even with relatively compact objects present in the CTR dataset. In order to further test the CTR-Base method, we evaluated it on the POT-210 [18] dataset, converting the ground -object corners -to segmentation (not modeling occlusions). The mean IoU (mIoU) is 0.81, showing that our method generalizes to POT-210 well.…”