import random from itertools import combinations import numpy as np from osdsynth.processor.pointcloud import calculate_distances_between_point_clouds, human_like_distance # from osdsynth.processor.prompt_template import * from osdsynth.processor.prompt_utils import * # from osdsynth.processor.prompt_spatitalbench_template import * from osdsynth.processor.prompt import * def camera_front_camera_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 15 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([0,0,-1])), -1.0, 1.0)) is_front = angle_rad < max_angle / 180 * np.pi check = is_front question_template = f"Does the camera face the front of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_back_camera_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 15 angle_rad = np.arccos(np.clip(np.dot(-A_rotation_matrix.T[0], np.array([0,0,-1])), -1.0, 1.0)) is_back = angle_rad < max_angle / 180 * np.pi check = is_back question_template = f"Does the camera face the back of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_left_camera_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 30 angle_rad = np.arccos(np.clip(np.dot(-A_rotation_matrix.T[2], np.array([0,0,-1])), -1.0, 1.0)) is_left = angle_rad < max_angle / 180 * np.pi check = is_left question_template = f"Does the camera face the left of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(60-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_right_camera_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 30 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[2], np.array([0,0,-1])), -1.0, 1.0)) is_right = angle_rad < max_angle / 180 * np.pi check = is_right question_template = f"Does the camera face the right of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(60-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_front_object_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 15 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([0,0,-1])), -1.0, 1.0)) is_front = angle_rad < max_angle / 180 * np.pi check = is_front question_template = f"Does the camera face the front of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_back_object_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 15 angle_rad = np.arccos(np.clip(np.dot(-A_rotation_matrix.T[0], np.array([0,0,-1])), -1.0, 1.0)) is_back = angle_rad < max_angle / 180 * np.pi check = is_back question_template = f"Does the camera face the back of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_left_object_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 30 angle_rad = np.arccos(np.clip(np.dot(-A_rotation_matrix.T[2], np.array([0,0,-1])), -1.0, 1.0)) is_left = angle_rad < max_angle / 180 * np.pi check = is_left question_template = f"Does the camera face the left of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(60-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def camera_right_object_center(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] A_rotation_matrix = A["rotation_matrix"] max_angle = 30 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[2], np.array([0,0,-1])), -1.0, 1.0)) is_right = angle_rad < max_angle / 180 * np.pi check = is_right question_template = f"Does the camera face the right of [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi-max_angle)/(60-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def object_side_by_side_same_direction(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_P_A = B_rotation_matrix.T @ (A_pos - B_pos) # 在B物体参考系下，A物体的位置 A_rotation_matrix = B_rotation_matrix.T @ A["rotation_matrix"] max_angle = 30 side_by_side_radius = np.abs(np.arctan(B_P_A[2]/ B_P_A[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) / 180 * np.pi # 比较X轴的夹角 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_same_orientation = angle_rad < max_angle / 180 * np.pi check = is_same_orientation and is_side_by_side question_template = f"Is [A] and [B] side by side, facing the same direction?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(angle_rad - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1 < 0 or w2 < 0 else w1 * w2 score = 0 if score < 0 else score return question, answer, check, score def object_side_by_side_opposite_direction(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_P_A = B_rotation_matrix.T @ (A_pos - B_pos) # 在B物体参考系下，A物体的位置 A_rotation_matrix = B_rotation_matrix.T @ A["rotation_matrix"] max_angle = 30 side_by_side_radius = np.abs(np.arctan(B_P_A[2]/ B_P_A[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) / 180 * np.pi angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([-1,0,0])), -1.0, 1.0)) is_opposite_orientation = angle_rad < max_angle / 180 * np.pi check = is_opposite_orientation and is_side_by_side question_template = f"Is [A] and [B] side by side, facing the opposite direction?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(angle_rad - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1 < 0 or w2 < 0 else w1 * w2 score = 0 if score < 0 else score return question, answer, check, score def object_face_to_face(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_P_A = B_rotation_matrix.T @ (A_pos - B_pos) # 在B物体参考系下，A物体的位置 A_rotation_matrix = B_rotation_matrix.T @ A["rotation_matrix"] max_angle = 30 face_to_face_radius = np.abs(np.arctan(B_P_A[2]/ B_P_A[0])) is_line = B_P_A[0] > 0 and face_to_face_radius < max_angle * np.pi# 在一条线上，且A在B的前面 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], [-1,0,0]), -1.0, 1.0)) is_opposite_orientation = angle_rad < max_angle / 180 * np.pi check = is_opposite_orientation and is_line question_template = f"Is [A] and [B] face to face?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 if B_P_A[0] > 0 else -1 w2 = 1 - np.abs(face_to_face_radius - max_angle * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(angle_rad - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1 < 0 or w2 < 0 or w3 < 0 else w1 * w2 * w3 score = 0 if score < 0 else score return question, answer, check, score def object_back_to_back(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_P_A = B_rotation_matrix.T @ (A_pos - B_pos) # 在B物体参考系下，A物体的位置 A_rotation_matrix = B_rotation_matrix.T @ A["rotation_matrix"] max_angle = 30 face_to_face_radius = np.abs(np.arctan(B_P_A[2]/ B_P_A[0])) is_line = B_P_A[0] < 0 and face_to_face_radius < max_angle / 180 * np.pi# 在一条线上，且A在B的前面 angle_rad = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], [-1,0,0]), -1.0, 1.0)) is_opposite_orientation = angle_rad < max_angle / 180 * np.pi check = is_opposite_orientation and is_line question_template = f"Is [A] and [B] back to back?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 if B_P_A[0] < 0 else -1 w2 = 1 - np.abs(face_to_face_radius - max_angle * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(angle_rad - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1 < 0 or w2 < 0 or w3 < 0 else w1 * w2 * w3 score = 0 if score < 0 else score return question, answer, check, score def object_front(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] A_rotation_matrix = A["rotation_matrix"] A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A物体参考系下，B物体的位置 max_angle = 15 A_P_B_direcetion = A_P_B / np.linalg.norm(A_P_B) angle_rad = np.arccos(np.clip(np.dot(A_P_B_direcetion, np.array([1,0,0])), -1.0, 1.0)) B_is_in_front_A = A_P_B[0] > 0 and angle_rad < max_angle / 180 * np.pi# 在一条线上，且A在B的前面 check = B_is_in_front_A question_template = f"Is [B] in front of [A], from the view of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi - max_angle) / (45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def object_back(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] A_rotation_matrix = A["rotation_matrix"] A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A物体参考系下，B物体的位置 max_angle = 15 A_P_B_direcetion = A_P_B / np.linalg.norm(A_P_B) angle_rad = np.arccos(np.clip(np.dot(A_P_B_direcetion, np.array([-1,0,0])), -1.0, 1.0)) B_is_in_back_A = A_P_B[0] < 0 and angle_rad < max_angle / 180 * np.pi# 在一条线上，且A在B的前面 check = B_is_in_back_A question_template = f"Is [B] in back of [A], from the view of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi - max_angle) / (45-max_angle)) score = 0 if score < 0 else score return question, answer, check, score def object_left(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] A_rotation_matrix = A["rotation_matrix"] A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A物体参考系下，B物体的位置 max_angle = 30 A_P_B_direcetion = A_P_B / np.linalg.norm(A_P_B) angle_rad = np.arccos(np.clip(np.dot(A_P_B_direcetion, np.array([0,0,-1])), -1.0, 1.0)) B_is_in_left_A = A_P_B[2] < 0 and angle_rad < max_angle / 180 * np.pi# 在一条线上，且A在B的前面 check = B_is_in_left_A question_template = f"Is [B] on the left of [A], from the view of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi - max_angle) / (60-max_angle)) score = 0 if score < 0 or A_P_B[2] > 0 else score return question, answer, check, score def object_right(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # A_rotation_matrix = A["rotation_matrix"] A_rotation_matrix = A["rotation_matrix"] A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A物体参考系下，B物体的位置 max_angle = 30 A_P_B_direcetion = A_P_B / np.linalg.norm(A_P_B) angle_rad = np.arccos(np.clip(np.dot(A_P_B_direcetion, np.array([0,0,1])), -1.0, 1.0)) B_is_in_right_A = A_P_B[2] > 0 and angle_rad < max_angle / 180 * np.pi# 在一条线上，且A在B的前面 check = B_is_in_right_A question_template = f"Is [B] on the right of [A], from the view of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs((angle_rad*180/np.pi - max_angle) / (60-max_angle)) score = 0 if score < 0 or A_P_B[2] < 0 else score return question, answer, check, score def camera_two_objects_closer(A,B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance_a = np.linalg.norm(A_pos) distance_b = np.linalg.norm(B_pos) check = distance_a < distance_b question_template = f"Is [A] closer to the camera than [B]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 1 if check else 0 return question, answer, check, score def camera_two_objects_farther(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance_a = np.linalg.norm(A_pos) distance_b = np.linalg.norm(B_pos) check = distance_a > distance_b question_template = f"Is [A] farther to the camera than [B]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 1 if check else 0 return question, answer, check, score def camera_two_objects_left(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] check = B_pos[0] - A_pos[0] > 0 question_template = f"Is [A] on the left of [B], from the view of the camera?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 1 if check else 0 return question, answer, check, score def camera_two_objects_right(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] check = B_pos[0] - A_pos[0] < 0 question_template = f"Is [A] on the right of [B], from the view of the camera?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 1 if check else 0 return question, answer, check, score def object_apart_0_5meter(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance = np.linalg.norm(A_pos - B_pos) delta = 1.0/3 gt_distance = 0.5 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] apart from [B] about 0.5 meter?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_apart_1meter(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance = np.linalg.norm(A_pos - B_pos) delta = 1.0/3 gt_distance = 1 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] apart from [B] about 1 meter?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_apart_1_5meter(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance = np.linalg.norm(A_pos - B_pos) delta = 1.0/3 gt_distance = 1.5 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] apart from [B] about 0.5 meter?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_apart_2meter(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] # 计算距离 distance = np.linalg.norm(A_pos - B_pos) delta = 1.0/3 gt_distance = 2 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] apart from [B] about 2 meters?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def camera_1meter_away(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] delta = 1.0/3 gt_distance = 1 distance = np.linalg.norm(A_pos) check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is the camera about 1 meter away from [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def camera_2meter_away(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] delta = 1.0/3 gt_distance = 2 distance = np.linalg.norm(A_pos) check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is the camera about 2 meter away from [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def camera_3meter_away(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] delta = 0.3 gt_distance = 3 distance = np.linalg.norm(A_pos) check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is the camera about 3 meter away from [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def camera_4meter_away(A): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] delta = 0.2 gt_distance = 4 distance = np.linalg.norm(A_pos) check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is the camera about 4 meter away from [A]?" question = question_template.replace("[A]", A_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_bigger_than1_2(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 计算距离 A_rotation_matrix = A["rotation_matrix"] theta_A = np.arctan2(A_rotation_matrix.T[0][2], A_rotation_matrix.T[0][0]) A_center = A["pcd"].get_center() R = A["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_A)) A["pcd"] = A["pcd"].rotate(R) A_length = A["pcd"].get_axis_aligned_bounding_box().get_extent()[0] A_height = A["pcd"].get_axis_aligned_bounding_box().get_extent()[1] A_width = A["pcd"].get_axis_aligned_bounding_box().get_extent()[2] A_volume = A_length * A_height * A_width B_rotation_matrix = B["rotation_matrix"] theta_B = np.arctan2(B_rotation_matrix.T[0][2], B_rotation_matrix.T[0][0]) B_center = B["pcd"].get_center() R = B["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_B)) B["pcd"] = B["pcd"].rotate(R) B_length = B["pcd"].get_axis_aligned_bounding_box().get_extent()[0] B_height = B["pcd"].get_axis_aligned_bounding_box().get_extent()[1] B_width = B["pcd"].get_axis_aligned_bounding_box().get_extent()[2] B_volume = B_length * B_height * B_width if A_volume > B_volume: distance = A_volume / B_volume else: distance = B_volume / A_volume delta = 1.0/3 gt_distance = 1.2 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] bigger than [B] about 0.2 times?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_higher_20cm(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 计算距离 A_height = A["pcd"].get_axis_aligned_bounding_box().get_extent()[1] B_height = B["pcd"].get_axis_aligned_bounding_box().get_extent()[1] distance = np.abs(A_height-B_height) delta = 1.0/3 gt_distance = 0.2 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] higher 20cm than [B]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_longer_50cm(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 计算距离 A_rotation_matrix = A["rotation_matrix"] theta_A = np.arctan2(A_rotation_matrix.T[0][2], A_rotation_matrix.T[0][0]) A_center = A["pcd"].get_center() R = A["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_A)) A["pcd"] = A["pcd"].rotate(R) A_length = A["pcd"].get_axis_aligned_bounding_box().get_extent()[0] B_rotation_matrix = B["rotation_matrix"] theta_B = np.arctan2(B_rotation_matrix.T[0][2], B_rotation_matrix.T[0][0]) B_center = B["pcd"].get_center() R = B["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_B)) B["pcd"] = B["pcd"].rotate(R) B_length = B["pcd"].get_axis_aligned_bounding_box().get_extent()[0] distance = np.abs(A_length-B_length) delta = 1.0/3 gt_distance = 0.5 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] longer 50cm than [B]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def object_wider_30cm(A, B): A_desc, A_cloud = A["caption"], A["pcd"] B_desc, B_cloud = B["caption"], B["pcd"] A_desc, B_desc = A_desc.lower(), B_desc.lower() # 计算距离 A_rotation_matrix = A["rotation_matrix"] theta_A = np.arctan2(A_rotation_matrix.T[0][2], A_rotation_matrix.T[0][0]) A_center = A["pcd"].get_center() R = A["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_A)) A["pcd"] = A["pcd"].rotate(R) A_width = A["pcd"].get_axis_aligned_bounding_box().get_extent()[2] B_rotation_matrix = B["rotation_matrix"] theta_B = np.arctan2(B_rotation_matrix.T[0][2], B_rotation_matrix.T[0][0]) B_center = B["pcd"].get_center() R = B["pcd"].get_rotation_matrix_from_xyz((0, 0, theta_B)) B["pcd"] = B["pcd"].rotate(R) B_width = B["pcd"].get_axis_aligned_bounding_box().get_extent()[2] distance = np.abs(A_width-B_width) delta = 1.0/3 gt_distance = 0.3 check = (1-delta)*gt_distance < distance and distance < (1+delta)*gt_distance question_template = f"Is [A] wider 30cm than [B]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: score = 1 - 1*np.abs(((distance - gt_distance) / gt_distance)- delta)/delta score = 0 if score < 0 else score return question, answer, check, score def side_by_side_front(A, B): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() B_desc, B_cloud = B["caption"], B["pcd"] B_desc = B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_rotation_matrix = A_rotation_matrix.T @ B_rotation_matrix # 在A的坐标系下，B的旋转矩阵 max_angle = 30 A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A的坐标系下，B相对于A的位置 side_by_side_radius = np.abs(np.arctan(A_P_B[2]/ A_P_B[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) * np.pi / 180 same_direction_radius = np.arccos(np.clip(np.dot(B_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_same_direction = same_direction_radius < max_angle * np.pi / 180 # 30度的阈值 front_radius = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([0,0,-1])), -1.0, 1.0)) is_front = front_radius < max_angle * np.pi / 180 # 30度的阈值 check = is_side_by_side and is_same_direction and is_front question_template = f"Is [A] and [B] side-by-side and same-orientation with viewed from the front of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(same_direction_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(front_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1<0 or w2<0 or w3<0 else w1 * w2 * w3 return question, answer, check, score def side_by_side_left(A, B): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() B_desc, B_cloud = B["caption"], B["pcd"] B_desc = B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_rotation_matrix = A_rotation_matrix.T @ B_rotation_matrix # 在A的坐标系下，B的旋转矩阵 max_angle = 30 A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A的坐标系下，B相对于A的位置 side_by_side_radius = np.abs(np.arctan(A_P_B[2]/ A_P_B[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) * np.pi / 180 same_direction_radius = np.arccos(np.clip(np.dot(B_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_same_direction = same_direction_radius < max_angle * np.pi / 180 # 30度的阈值 left_radius = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([-1,0,0])), -1.0, 1.0)) is_left = left_radius < max_angle * np.pi / 180 # 30度的阈值 check = is_side_by_side and is_same_direction and is_left question_template = f"Is [A] and [B] side-by-side and same-orientation with viewed from the left of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(same_direction_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(left_radius - max_angle / 180 * np.pi) / (np.pi / 6) # 30度的阈值 score = 0 if w1<0 or w2<0 or w3<0 else w1 * w2 * w3 return question, answer, check, score def side_by_side_right(A, B): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() B_desc, B_cloud = B["caption"], B["pcd"] B_desc = B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_rotation_matrix = A_rotation_matrix.T @ B_rotation_matrix # 在A的坐标系下，B的旋转矩阵 max_angle = 30 A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A的坐标系下，B相对于A的位置 side_by_side_radius = np.abs(np.arctan(A_P_B[2]/ A_P_B[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) * np.pi / 180 same_direction_radius = np.arccos(np.clip(np.dot(B_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_same_direction = same_direction_radius < max_angle * np.pi / 180 # 30度的阈值 right_radius = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_right = right_radius < max_angle * np.pi / 180 # 30度的阈值 check = is_side_by_side and is_same_direction and is_right question_template = f"Is [A] and [B] side-by-side and same-orientation with viewed from the right of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(same_direction_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(right_radius - max_angle / 180 * np.pi) / (np.pi / 6) # 30度的阈值 score = 0 if w1<0 or w2<0 or w3<0 else w1 * w2 * w3 return question, answer, check, score def side_by_side_back(A, B): A_desc, A_cloud = A["caption"], A["pcd"] A_desc = A_desc.lower() B_desc, B_cloud = B["caption"], B["pcd"] B_desc = B_desc.lower() # 从PyTorch3D的坐标系转换到OpenCV的坐标系 A_pos = A_cloud.get_center() A_pos[0] = -A_pos[0]; A_pos[1] = -A_pos[1] B_pos = B_cloud.get_center() B_pos[0] = -B_pos[0]; B_pos[1] = -B_pos[1] A_rotation_matrix = A["rotation_matrix"] B_rotation_matrix = B["rotation_matrix"] B_rotation_matrix = A_rotation_matrix.T @ B_rotation_matrix # 在A的坐标系下，B的旋转矩阵 max_angle = 30 A_P_B = A_rotation_matrix.T @ (B_pos - A_pos) # 在A的坐标系下，B相对于A的位置 side_by_side_radius = np.abs(np.arctan(A_P_B[2]/ A_P_B[0])) is_side_by_side = side_by_side_radius > (90 - max_angle) * np.pi / 180 same_direction_radius = np.arccos(np.clip(np.dot(B_rotation_matrix.T[0], np.array([1,0,0])), -1.0, 1.0)) is_same_direction = same_direction_radius < max_angle * np.pi / 180 # 30度的阈值 back_radius = np.arccos(np.clip(np.dot(A_rotation_matrix.T[0], np.array([0,0,1])), -1.0, 1.0)) is_back = back_radius < max_angle * np.pi / 180 # 30度的阈值 check = is_side_by_side and is_same_direction and is_back question_template = f"Is [A] and [B] side-by-side and same-orientation with viewed from the back of [A]?" question = question_template.replace("[A]", A_desc).replace("[B]", B_desc) answer = "Yes" if check else "No" score = 0 if check: score = 1 else: w1 = 1 - np.abs(side_by_side_radius - (90 - max_angle) / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w2 = 1 - np.abs(same_direction_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 w3 = 1 - np.abs(back_radius - max_angle / 180 * np.pi) / (np.pi / 12) # 15度的阈值 score = 0 if w1<0 or w2<0 or w3<0 else w1 * w2 * w3 return question, answer, check, score