import random from itertools import combinations import numpy as np from osdsynth.processor.prompt_utils import * from osdsynth.processor.prompt_T2Ibench import * from osdsynth.processor.prompt_ImageEditbench import * from osdsynth.processor.prompt_CR import * class T2IPromptGenerator: def __init__(self, cfg, logger, device): """Initialize the class.""" self.cfg = cfg self.logger = logger self.device = device self.vis = True def evaluate_predicates_on_pairs(self, detections, n_conv=3, spatial_choice=-1): # 全部是SpatialBench的prompt all_prompt_variants = [ camera_front_camera_center, # 0 camera_back_camera_center, # 1 camera_left_camera_center, # 2 camera_right_camera_center, # 3 camera_front_object_center, # 4 camera_back_object_center, # 5 camera_left_object_center, # 6 camera_right_object_center, # 7 object_side_by_side_same_direction, # 8 object_side_by_side_opposite_direction, # 9 object_face_to_face, # 10 object_back_to_back, # 11 object_front, # 12 object_back, # 13 object_left, # 14 object_right, # 15 camera_two_objects_closer, # 16 camera_two_objects_farther, # 17 camera_two_objects_left, # 18 camera_two_objects_right, # 19 object_apart_0_5meter, # 20 object_apart_1meter, # 21 object_apart_1_5meter, # 22 object_apart_2meter, # 23 camera_1meter_away, # 24 camera_2meter_away, # 25 camera_3meter_away, # 26 camera_4meter_away, # 27 object_bigger_than1_2, # 28 object_higher_20cm, # 29 object_longer_50cm, # 30 object_wider_30cm, # 31 side_by_side_front, # 32 side_by_side_left, # 33 side_by_side_right, # 34 side_by_side_back, # 35 ] if spatial_choice != -1: all_prompt_variants = [all_prompt_variants[spatial_choice]] else: raise ValueError("spatial_choice must not be -1 for T2Ibench") if spatial_choice not in [0,1,2,3,4,5,6,7,24,25,26,27]: all_combinations = list(combinations(range(len(detections)), 2)) random.shuffle(all_combinations) selected_combinations = all_combinations[:3] object_pairs = [(detections[i], detections[j]) for i, j in selected_combinations] results = [] correct = 0 for A, B in object_pairs: all_prompt_variants = [item for item in all_prompt_variants] # selected_predicates_choices = random.sample(all_prompt_variants, n_conv) selected_predicates_choices = random.sample(all_prompt_variants, len(all_prompt_variants)) for prompt_func in selected_predicates_choices: res = prompt_func(A, B) results.append((res, A, B, prompt_func.__name__)) correct = correct + 1 if res[2] else correct score = res[3] return results, correct, score else: A = detections[0] results = [] correct = 0 all_prompt_variants = [item for item in all_prompt_variants] selected_predicates_choices = random.sample(all_prompt_variants, len(all_prompt_variants)) for prompt_func in selected_predicates_choices: res = prompt_func(A) results.append((res, A, prompt_func.__name__)) correct = correct + 1 if res[2] else correct score = res[3] return results, correct, score class ImageEditPromptGenerator: def __init__(self, cfg, logger, device): """Initialize the class.""" self.cfg = cfg self.logger = logger self.device = device self.vis = True def evaluate_predicates_on_pairs(self, detections, n_conv=3, spatial_choice=-1): # 全部是SpatialBench的prompt all_prompt_variants = [ camera_to_front_camera_center, # 0 camera_to_left_camera_center, # 1 camera_to_right_camera_center, # 2 camera_to_back_camera_center, # 3 camera_to_front_object_center, # 4 camera_to_left_object_center, # 5 camera_to_right_object_center, # 6 camera_to_back_object_center, # 7 object_insert_side_by_side_same_orientation, # 8 object_insert_side_by_side_opposite_orientation, # 9 object_insert_face_to_face, # 10 object_insert_back_to_back, # 11 object_insert_front_object_center, # 12 object_insert_left_object_center, # 13 object_insert_right_object_center, # 14 object_insert_behind_object_center, # 15 object_insert_front_camera_center, # 16 object_insert_left_camera_center, # 17 object_insert_right_camera_center, # 18 object_insert_behind_camera_center, # 19 objectmove_close_1meter, # 20 objectmove_far_1meter, # 21 objectmove_left_1meter, # 22 objectmove_right_1meter, # 23 camera_forward_1meter, # 24 camera_leftward_1meter, # 25 camera_rightward_1meter, # 26 camera_backward_1meter, # 27 object_make_12bigger, # 28 object_make_20cm_higher, # 29 object_make_50cm_longer, # 30 object_make_40cm_wider, # 31 ] if spatial_choice != -1: all_prompt_variants = [all_prompt_variants[spatial_choice]] else: raise ValueError("spatial_choice must not be -1 for T2Ibench") if spatial_choice not in [0,1,2,3,4,5,6,7,20,21,22,23,24,25,26,27,28,29,30,31]: object_pairs = [(detections[0], detections[1])] results = [] correct = 0 for A, B in object_pairs: all_prompt_variants = [item for item in all_prompt_variants] # selected_predicates_choices = random.sample(all_prompt_variants, n_conv) selected_predicates_choices = random.sample(all_prompt_variants, len(all_prompt_variants)) for prompt_func in selected_predicates_choices: res = prompt_func(A, B) results.append((res, A, B, prompt_func.__name__)) correct = correct + 1 if res[2] else correct score = res[3] return results, correct, score else: A = detections[0] results = [] correct = 0 all_prompt_variants = [item for item in all_prompt_variants] selected_predicates_choices = random.sample(all_prompt_variants, len(all_prompt_variants)) for prompt_func in selected_predicates_choices: res = prompt_func(A) results.append((res, A, prompt_func.__name__)) correct = correct + 1 if res[2] else correct score = res[3] return results, correct, score class CRPromptGenerator: def __init__(self, cfg, logger, device): """Initialize the class.""" self.cfg = cfg self.logger = logger self.device = device self.vis = True def evaluate_predicates_on_pairs(self, detections, is_three=False, spatial_choice=-1): # 全部是SpatialBench的prompt all_prompt_variants_two = [ CR_two_front, CR_two_back, CR_two_left, CR_two_right, ] all_prompt_variants_three = [ CR_three_front, CR_three_back, CR_three_left, CR_three_right, ] A = detections[0] B = detections[1] if is_three: C = detections[2] all_prompt_variants = all_prompt_variants_three else: C = None all_prompt_variants = all_prompt_variants_two if spatial_choice != -1: all_prompt_variants = [all_prompt_variants[spatial_choice]] else: raise ValueError("spatial_choice must not be -1 for T2Ibench") results = [] correct = 0 all_prompt_variants = [item for item in all_prompt_variants] # selected_predicates_choices = random.sample(all_prompt_variants, n_conv) selected_predicates_choices = random.sample(all_prompt_variants, len(all_prompt_variants)) prompt_func = selected_predicates_choices[0] if is_three: res = prompt_func(A, B, C) results.append((res, A, B, C, prompt_func.__name__)) else: res = prompt_func(A, B) results.append((res, A, B, prompt_func.__name__)) correct = correct + 1 if res[2] else correct score = res[3] return results, correct, score