LookOut | Tracker Comparison

Here we visualize tracker performance across different scenes covering single and multiple (two and three) actors and occlusion difficulty.

Unlike IRL Tracker Comparison where we compare tracker performance in our system end-to-end on a real-life sequence filmed multiple times, here we compare tracker performance after the scene had been filmed. The crucial difference is that while a tracker distraction will move the gimbal in the wrong direction and ruin a scene, here the camera view direction is fixed so trackers still have the benefit of a distraction moving out of frame and the target staying in frame.

Each tracker is instantiated with a single bounding box at the beginning of the sequence and judgement starts immediately, so no trackers that rely on time for building a feature history - including Ours - get an edge. We label each moment a tracker gets distracted in the videos with a big red circle along with a count of the number of frames the tracker has reported tracking on an incorrect target. For Market and TwoPeople we label these distraction moments by leveraging groundtruth bounding-box labels of each actor, which means single frame slips during graceful track loss reporting are harshly scrutinized. For the rest of the videos, we manually label frames with distraction for each tracker.

We report FPS counts for each tracker (top-left in each panel) to showcase speed during dense scenes and how speed scales with multiple actors. Note: here we color code tracker status differently to the rest of supplemental website. Orange () indicates a tracker is searching for a target, Purple () indicates a track is in the recovery stage (ours only), and Green () indicates a tracker thinks it is tracking the target.

All videos here, with the exception of NY, were filmed during LookOut's development so the system's maturity in these videos is variable.