While it may seem more logical to label your 2D objects first, at Segments.ai, we advise you to start with 3D labels first.
2D bounding boxes are fast to annotate, but you have to draw them in every single frame. A single traffic sign can mean 300 2D bounding boxes if you have three different cameras — a labeling task that can cost hours of your time.
When you start with 3D though, you can leverage the power of scale. The same traffic sign only requires a single 3D cuboid, making it 100 times more efficient than the 2D bounding boxes. You can then project that cuboid onto the 2D data and automatically fill in the labels.