# object_direction_map.yaml # # Maps (object facing direction, original question direction) → camera-centric direction. # # How to read this table: # Given the direction RoboBrain says the object is facing AND the direction keyword in the # original question (object-centric), look up the camera-centric replacement direction. # # ─── Facing keys (top-level) ────────────────────────────────────────────────────────────── # facing_left : the object faces toward camera-left (west in bird's-eye view) # facing_right : the object faces toward camera-right (east) # facing_toward_camera : the object faces the camera (south) # facing_away_from_camera: the object faces away from camera (north) # # ─── Original direction keys (per facing) ───────────────────────────────────────────────── # front : object-centric "in front of" (the direction the object's face points) # behind : object-centric "behind" (opposite to the face direction) # left : object-centric "left" (the object's own left side) # right : object-centric "right" (the object's own right side) # # ─── Values (camera-centric direction) ─────────────────────────────────────────────────── # left → question phrase "left of" # right → question phrase "right of" # in_front_of → question phrase "in front of" # behind → question phrase "behind" # # ─── Geometric derivation (bird's-eye view, camera at south looking north) ─────────────── # If object faces west (left): # forward=west=left, back=east=right, # obj-left = south = behind (camera), obj-right = north = in front of (camera) # If object faces east (right): # forward=east=right, back=west=left, # obj-left = north = in_front_of, obj-right = south = behind # If object faces south (toward camera): # forward=south=behind, back=north=in_front_of, # obj-left = east = right, obj-right = west = left # If object faces north (away from camera): # forward=north=in_front_of, back=south=behind, # obj-left = west = left, obj-right = east = right # # NOTE: These values are derived from standard 3-D geometry. Please verify them # against your data before running a full evaluation. facing_left: front: left behind: right left: behind right: in_front_of facing_right: front: right behind: left left: in_front_of right: behind facing_toward_camera: front: behind behind: in_front_of left: right right: left facing_away_from_camera: front: in_front_of behind: behind left: left right: right