Boxy Vehicle Detection in Large Images

September 2019

tl;dr: A large dataset with 3D-like labels from Bosch.

Overall impression

The author proposed to annotate cuboids with 2 plans, one axis-aligned bounding box (AABB) for rear side and the other trapezoid for the side. This annotation idea is brilliant. The authors did mention the upper front point is ambiguous to label.

Another big dataset with 3D-like label is BoxCars, but BoxCars is a surveillance dataset and the vehicle angle is different. (For example, in surveillance, we could see the top of the car in most images, but in almost none of the autonomous driving scenes).

Key ideas

Technical details