 Sequence
1 - Object Detections
This video shows our system's object detections using automatically
estimated groundplane constraints from Structure-from-Motion for our
first test sequence. For this sequence, we integrated the output of 5
single-view car detectors (plus 2 mirrored versions) for different car
viewpoints.
Download this video
This data set is available here.
|
 Sequence
1 - 3D Hypotheses
Results of the spacetime trajectory estimation. The video shows the
hypothesized 3D car locations estimated for every frame of the first
test sequence (using only detections from this frame and previous
frames). Fixed 3D bounding boxes are estimated for static cars;
trajectories and predicted motion directions are estimated for moving
cars. Depicted bounding boxes are color coded, with the intensity
corresponding to the system's confidence. It can be seen that as soon as
scene objects come into a range of 15-20m, 3D bounding boxes appear and
snap into place as soon as sufficient evidence has been accumulated.
Download this video
|
 Sequence
2 - Object Detections
This video shows our system's object detections using automatically
estimated groundplane constraints from Structure-from-Motion for our
second test sequence. For this sequence, we integrated the output of 5
single-view car detectors (plus 2 mirrored versions) for different car
viewpoints and an additional pedestrian detector.
Download this video
|
 Sequence
2 - Spacetime Trajectories
Visualization of the spacetime volume of detections and estimated
trajectories. Red dots indicate car detections; blue dots show
pedestrian detections. As our vehicle is driving through the street, new
detections come in and the 3D object locations and trajectories are
reestimated for every frame.
Download this video
|
 Sequence
2 - 3D Hypotheses
Detected object locations and estimated trajectories for the second
test sequence (again bounding boxes are color coded to indicate the
system's confidence in the displayed hypotheses). This very challenging
sequence has been recorded at only 3 fps, making it very difficult to
obtain any trajectories; consequently, it is not always possible to
filter out false positives from false detections at this low framerate.
Still, false positives typically get only low confidences and quickly
fade out as they fail to get continuous support.
Download this video
|