3d tracking benchmarking about 3d-vehicle-tracking HOT 2 CLOSED

Kitti training video 0001

You can clearly see the mismatch between dotted ground truth and solid predictions.

from 3d-vehicle-tracking.

Hi @bilalsattar, thanks for the figure and great question.

I think it's reasonable. Our work using a monocular camera as the visual input (per frame) to estimate the 3D bounding boxes. We are doing great for near vehicles. On the other hand, even a stereo camera setting can have a hard time in the 40-80m range. It would definitely produce reasonable errors for a monocular setting.

We found out that our 3D tracking pipeline helps 2D association and vice versa. Even today, our work stayed state-of-the-art results (top 10) on the KITTI benchmark.

Lastly, we encourage you to take a look at our new work using quasi-dense representation learning with more robust 3D estimation on bigger datasets (e.g., nuScenes, Waymo). Please have a look here.

from 3d-vehicle-tracking.