Object Detection Device Using Direction-Aware 2D Bounding Box Transformation

The patent transitions from using only 2D bounding box center positions to incorporating 3D direction information (orientation angle) to accurately represent object position. By adding the dimension of directional information, the system can distinguish between objects at different orientations even when their 2D projections overlap, thereby improving position estimation accuracy without excessive complexity increase.

The patent segments the object representation into distinct components: 2D bounding box position, object direction (orientation angle), and 3D size information. This segmentation allows each component to be processed and transformed independently, maintaining computational efficiency while improving overall position estimation accuracy through the combination of these segmented information elements.

The patent uses 2D bounding box information as a copy or proxy for the more complex 3D bounding box representation. Instead of training a neural network to directly output 3D bounding boxes requiring expensive 3D annotation data, the system uses the simpler 2D bounding box annotations to infer object position and direction, thereby reducing the quantity and cost of required annotation data while maintaining estimation accuracy.

The patent employs inexpensive 2D bounding box annotations as a substitute for expensive 3D bounding box annotations. The 2D bounding box, which can be annotated at low cost, serves as a disposable intermediate representation that suffices for the detection task, eliminating the need for costly 3D annotation data while still enabling accurate object position and size estimation.