A9 dataset for monocular 3D object detection

[johannes-tum]

Hi, I would be interest in using your dataset for monocular 3D object detection. However, I noticed that unfortunately you provide mostly bounding box corner points projected onto the image. That's what I found:

a9_dataset_r00_s00: ca. 800 images, but only projected bboxes a9_dataset_r00_s01: ca. 200 images, projected bboxes + location a9_dataset_r00_s02: ca. 60 images, projected bboxes + location a9_dataset_r01_s01: ca. 1500 images, but only projected bboxes a9_dataset_r01_s01: ca. 1500 images, but only projected bboxes a9_dataset_r01_s03: ca. 3000 images, but only projected bboxes

Would you also be able to provide full 3D informationframe for all these cases? With that I would mean:

• 3D dimension (height, width, length)
• 3D location in camera frame (X, Y, Z)
• 3D rotation. Ideally the rotation matrix in camera frame.

[walzimmer]

Hi Johannes,

for monocular 3D object detection please use the sequences below. They contain 3D bounding boxes (3D location, 3D dimensions and yaw rotation):

TUM Traffic Highway Dataset: Sequence a9_dataset_r00_s01 and a9_dataset_r00_s02.

• contains 642 labeled images with 11,355 labeled 3D bounding boxes

TUM Traffic Intersection Dataset: Sequence a9_dataset_r02_s01, a9_dataset_r02_s02, a9_dataset_r02_s03, and a9_dataset_r02_s04.

• contains 4,800 labeled images with 62,396 labeled 3D bounding boxes

Best, Walter