Update README.md

README.md

@@ -140,7 +140,7 @@ We include a tight 9-DOF bounding box for each semantic instance in `ai_VVV_NNN/
 
 `metadata_semantic_instance_bounding_box_object_aligned_2d_positions.hdf5` contains an Nx3 array of bounding box center positions, where N is the number of semantic instances, and each position is stored in [x,y,z] order. These positions are specified in asset coordinates.
 
-We compute each bounding box's rotation matrix according to the following algorithm. We always set the positive z-axis of our rotation matrix to point up, i.e., to align with the world-space gravity vector. We then compute a 2D minimum-area bounding box in the world-space xy-plane. Once we have computed our minimum-area bounding box, we have 4 possible choices for the positive x-axis of our rotation matrix. To make this choice, we consider the vector from the bounding box's geometric center to the center-of-mass of the points used to compute the bounding box. We choose the direction (among our 4 possible choices) that most closely aligns with this vector as the positive x-axis of our rotation matrix. Finally, we set the positive y-axis to be our positive x-axis rotated by +90 degrees in the world space xy-plane (i.e., so our rotation matrix will have a determinant of 1). This algorithm encourages that similar objects with semantically similar orientations will be assigned rotation matrices that are similar (i.e., the difference of their rotation matrices will have a small matrix norm).
+We compute each bounding box's rotation matrix according to the following algorithm. We always set the positive z-axis of our rotation matrix to point up, i.e., to align with the world-space gravity vector. We then compute a 2D minimum-area bounding box in the world-space xy-plane. Once we have computed our minimum-area bounding box, we have 4 possible choices for the positive x-axis of our rotation matrix. To make this choice, we consider the vector from the bounding box's geometric center to the center-of-mass of the points used to compute the bounding box. We choose the direction (among our 4 possible choices) that most closely aligns with this vector as the positive x-axis of our rotation matrix. Finally, we set the positive y-axis to be our positive x-axis rotated by +90 degrees in the world-space xy-plane (i.e., so our rotation matrix will have a determinant of 1). This algorithm encourages that similar objects with semantically similar orientations will be assigned rotation matrices that are similar (i.e., the difference of their rotation matrices will have a small matrix norm).
 
 Our code can be used to compute other types of bounding boxes (e.g., axis-aligned in world-space, minimum-volume), but we don't include these other types of bounding boxes in our public release.