search

apple / ml-hypersim

Hypersim: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding
Other
1.71k stars 131 forks source link

generate 3D bounding boxes #48

Closed dferthvf closed 2 years ago

dferthvf commented 2 years ago

Hi, when I tried to use the script "scene_generate_images_bounding_box.py" to generate bounding boxes, I wondered what is the difference between bounding_box_type == "object_aligned_2d" and bounding_box_type == "object_aligned_3d".

mikeroberts3000 commented 2 years ago

Hi, great question. The object_aligned_2d method is the one described in the main README. Using this method, one of the bounding box's axes is always aligned with the world-space up vector (i.e., the gravity vector), and the other two axes are computed such that the bounding box has minimum area. The object_aligned_3d method computes the minimum-volume bounding box, and in general does not produce bounding boxes that are aligned with the world-space up vector.