add option to change IoU threshold

functions/binary_search.py

@@ -84,9 +84,9 @@ def trapezoidal_interpolation(lower_value, upper_value, lower_metric, upper_metr
     return lower_value
 
 
-def filter_table_and_get_metrics(args, tableGT, tablePred, threshold, i=0):
+def filter_table_and_get_metrics(args, tableGT, tablePred, threshold, i=0, iou_threshold=0.2):
     filtered_tableGT, filtered_tablePred = filter_by_proba_threshold(tableGT, tablePred, threshold)
-    image_level_df = compute_image_level_metrics(args, filtered_tableGT, filtered_tablePred)
+    image_level_df = compute_image_level_metrics(args, filtered_tableGT, filtered_tablePred, iou_threshold)
 
     if args.debug:
         image_tsv_path = os.path.join(args.output_path, f'image_level_{threshold}_iteration_{i}.tsv')
@@ -95,7 +95,7 @@ def filter_table_and_get_metrics(args, tableGT, tablePred, threshold, i=0):
     return calculate_average_metrics(image_level_df)
 
 
-def binary_search_for_threshold(args, target_metric, tableGT, tablePred, metric_type):
+def binary_search_for_threshold(args, target_metric, tableGT, tablePred, metric_type, iou_threshold=0.2):
     # Extract sorted unique probability values from predicted probabilities
     unique_probas = sorted(tablePred['predicted_proba'].unique())
     unique_probas.insert(0,0)
@@ -114,7 +114,8 @@ def binary_search_for_threshold(args, target_metric, tableGT, tablePred, metric_
         proba_threshold = unique_probas[mid]
 
         # Filter tables by current threshold, and calculate image-level and average metrics
-        mean_sensitivity, mean_precision, mean_f1_score, mean_fp_per_image = filter_table_and_get_metrics(args, tableGT, tablePred, proba_threshold, i)
+        mean_sensitivity, mean_precision, mean_f1_score, mean_fp_per_image = \
+            filter_table_and_get_metrics(args, tableGT, tablePred, proba_threshold, i, iou_threshold)
 
         # Select the metric of interest
         current_metric = {
@@ -131,7 +132,8 @@ def binary_search_for_threshold(args, target_metric, tableGT, tablePred, metric_
                     proba_threshold, mean_sensitivity, mean_precision, mean_f1_score, mean_fp_per_image
                 mid -= 1
                 proba_threshold = unique_probas[mid]
-                mean_sensitivity, mean_precision, mean_f1_score, mean_fp_per_image = filter_table_and_get_metrics(args, tableGT, tablePred, proba_threshold)
+                mean_sensitivity, mean_precision, mean_f1_score, mean_fp_per_image = \
+                    filter_table_and_get_metrics(args, tableGT, tablePred, proba_threshold, i, iou_threshold)
                 current_metric = {
                     'mean_fp_per_image': mean_fp_per_image,
                     'mean_sensitivity': mean_sensitivity,
@@ -161,19 +163,20 @@ def binary_search_for_threshold(args, target_metric, tableGT, tablePred, metric_
         return proba_threshold, mean_sensitivity, mean_precision, mean_f1_score 
 
 
-def single_binary_search(args, gt_path, pred_path, target_metric, metric_type, output_path):
+def single_binary_search(args, gt_path, pred_path, target_metric, metric_type, output_path, iou_threshold=0.2):
     # Load data
     tableGT, tablePred = load_data(gt_path, pred_path)
 
     # Binary search for optimal threshold based on target metric
     final_threshold, mean_sensitivity, mean_precision, mean_f1_score = binary_search_for_threshold(
-        args, target_metric, tableGT, tablePred, metric_type
+        args, target_metric, tableGT, tablePred, metric_type, iou_threshold
     )
 
     if args.debug:
         # Save the final image-level metrics to TSV
         filtered_tableGT_final, filtered_tablePred_final = filter_by_proba_threshold(tableGT, tablePred, final_threshold)
-        image_level_df_final = compute_image_level_metrics(args, filtered_tableGT_final, filtered_tablePred_final)
+        image_level_df_final = compute_image_level_metrics(args, filtered_tableGT_final, filtered_tablePred_final,
+                                                           iou_threshold)
         image_tsv_path = os.path.join(output_path, f'image_level_{metric_type}_{target_metric}_{final_threshold}.tsv')
         image_level_df_final.to_csv(image_tsv_path, sep='\t', index=False)
 

functions/froc_calculation.py

@@ -8,14 +8,14 @@ import matplotlib.pyplot as plt
 from .binary_search import single_binary_search
 
 
-def calculate_froc(args, gt_path, pred_path, target_metric_list, metric_type, output_path):
+def calculate_froc(args, gt_path, pred_path, target_metric_list, metric_type, output_path, iou_threshold=0.2):
     mean_sensitivities = []
     results = {}
 
     # Run binary search for each target metric and collect sensitivities
     for target_metric in target_metric_list:
         final_threshold, mean_sensitivity, mean_precision, mean_f1_score = single_binary_search(
-            args, gt_path, pred_path, target_metric, metric_type, output_path
+            args, gt_path, pred_path, target_metric, metric_type, output_path, iou_threshold
         )
         mean_sensitivities.append(mean_sensitivity)
         result_tmp = {
@@ -68,6 +68,7 @@ if __name__ == "__main__":
     parser.add_argument('-m', '--metric_type', required=False, type=str, default='mean_fp_per_image', help='Metrics to reach.')
     parser.add_argument('-tm', '--target_metric_list', required=False, type=list, default=[0.25, 0.5, 1, 2, 3], help='Value to reach for the selected target metrics.')
     parser.add_argument('-o', '--output_path', required=False, default='/dataEvalChallenge/tmp', help='Path to the output folder.')
+
     parser.add_argument('-debug', '--debug', action='store_true', help='Printing some intermediate information.')
     args = parser.parse_args()
 

main.py

@@ -6,7 +6,8 @@ from functions.froc_calculation import calculate_froc
 
 def main(args):
     generate_lesion_level_files(args, args.gt_path, args.pred_path, args.csv_path, args.iou_tresh, args.output_path)
-    calculate_froc(args, args.gt_tsv_path, args.pred_tsv_path, args.target_metric_list, args.metric_type, args.output_path)
+    calculate_froc(args, args.gt_tsv_path, args.pred_tsv_path, args.target_metric_list, args.metric_type,
+                   args.output_path, args.iou_tresh)
 
 
 if __name__ == "__main__":