Update

app.py

@@ -19,14 +19,14 @@ def compute_iou(row):
     yA = max(row["ymin_gt"], row["ymin_pred"])
     xB = min(row["xmax_gt"], row["xmax_pred"])
     yB = min(row["ymax_gt"], row["ymax_pred"])
-    
+
     inter_width = max(0, xB - xA)
     inter_height = max(0, yB - yA)
     inter_area = inter_width * inter_height
-    
+
     boxA_area = (row["xmax_gt"] - row["xmin_gt"]) * (row["ymax_gt"] - row["ymin_gt"])
     boxB_area = (row["xmax_pred"] - row["xmin_pred"]) * (row["ymax_pred"] - row["ymin_pred"])
-    
+
     union_area = boxA_area + boxB_area - inter_area
     iou = inter_area / union_area if union_area > 0 else 0
     return iou
@@ -34,8 +34,8 @@ def compute_iou(row):
 def evaluate_submission(file, submission_name, model_description):
     """
     Evaluate the submitted CSV file by comparing predicted bounding box coordinates and food weight
-    against the ground truth. Rows with a weight value of -1 in either ground truth or submission
-    are filtered out.
+    against the ground truth. We merge on both 'image_name' and 'class_id', because there can be
+    multiple items (classes) in the same image.
     """
     if file is None:
         return "❌ Please upload a file."
@@ -46,8 +46,8 @@ def evaluate_submission(file, submission_name, model_description):
         print("Submitted columns:", submission_df.columns.tolist())
     except Exception as e:
         return f"❌ File read failed: {e}"
-    
-    # Load hidden ground truth CSV from a private dataset.
+
+    # Load hidden ground truth CSV from a private dataset (adjust repo_id as needed).
     try:
         gt_path = hf_hub_download(
             repo_id="issai/ground-truth-food-eval",
@@ -58,74 +58,61 @@ def evaluate_submission(file, submission_name, model_description):
         gt_df = pd.read_csv(gt_path)
     except Exception as e:
         return "Error loading ground truth file: " + str(e)
-    
-    # --- Debug: Print shapes and samples from both DataFrames ---
+
+    # Debug: Print shapes and first few rows.
     print("Ground truth shape:", gt_df.shape)
     print("Ground truth sample:")
     print(gt_df.head())
-    
+
     print("Submission shape:", submission_df.shape)
     print("Submission sample:")
     print(submission_df.head())
-    # ---------------------------------------------------------
-    
-    # Merge on 'image_id'
+
+    # Merge on 'image_name' and 'class_id' so that multiple classes per image match up correctly.
     try:
-        df = pd.merge(gt_df, submission_df, on="image_id", suffixes=("_gt", "_pred"))
+        df = pd.merge(
+            gt_df, submission_df,
+            on=["image_name", "class_id"],
+            suffixes=("_gt", "_pred")
+        )
     except Exception as e:
         return f"❌ Error during merge: {e}"
-    
+
     print(f"✅ Received {len(submission_df)} predictions for evaluation!")
     print("Merged shape:", df.shape)
-    
+
     if df.empty:
-        return "No matching image_ids found between ground truth and submission."
-    
+        return "No matching rows found between ground truth and submission (image_name/class_id)."
+
     print("Merged columns:", df.columns.tolist())
     print("Merged sample data:")
-    print(df[["image_id", "xmin_gt", "xmin_pred", "weight_gt", "weight_pred"]].head())
-    
-    # For further debugging: compute differences for one coordinate and weight.
+    print(df[["image_name", "class_id", "xmin_gt", "xmin_pred", "weight_gt", "weight_pred"]].head())
+
+    # Debug differences in xmin and weight
     df["diff_xmin"] = abs(df["xmin_gt"] - df["xmin_pred"])
     df["diff_weight"] = abs(df["weight_gt"] - df["weight_pred"])
     print("Differences in 'xmin' (head):", df["diff_xmin"].head())
     print("Differences in 'weight' (head):", df["diff_weight"].head())
 
-    # Find rows where either diff_xmin or diff_weight is not zero.
-    nonzero_diff = df[(df["diff_xmin"] != 0) | (df["diff_weight"] != 0)]
-    if not nonzero_diff.empty:
-        print("Rows with nonzero differences:")
-        print(nonzero_diff)
-    else:
-        print("No nonzero differences found.")
-    
-    # Convert coordinate and weight columns to numeric.
-    coord_cols = ["xmin_gt", "ymin_gt", "xmax_gt", "ymax_gt",
-                  "xmin_pred", "ymin_pred", "xmax_pred", "ymax_pred"]
+    # Convert columns to numeric
+    coord_cols = [
+        "xmin_gt", "ymin_gt", "xmax_gt", "ymax_gt",
+        "xmin_pred", "ymin_pred", "xmax_pred", "ymax_pred"
+    ]
     for col in coord_cols:
         df[col] = pd.to_numeric(df[col], errors="coerce")
+
     df["weight_gt"] = pd.to_numeric(df["weight_gt"], errors="coerce")
     df["weight_pred"] = pd.to_numeric(df["weight_pred"], errors="coerce")
-    
-    # Filter out rows where either weight is -1 ---
-    num_before = df.shape[0]
-    df = df[(df["weight_gt"] != -1) & (df["weight_pred"] != -1)]
-    num_filtered = num_before - df.shape[0]
-    print(f"Filtered out {num_filtered} rows due to weight == -1")
-    
-    if df.empty:
-        return "No evaluable rows remain after filtering rows with weight == -1."
-    
-    # Compute IoU for each row.
+
+    # Compute IoU for each row
     df["iou"] = df.apply(compute_iou, axis=1)
-    
-    # Compute mean absolute error for food weight predictions.
+
+    # Compute mean absolute error for weight
     mean_weight_error = mean_absolute_error(df["weight_gt"], df["weight_pred"])
-    
-    # Calculate mean IoU.
     mean_iou = df["iou"].mean()
-    
-    # Save the evaluation result to the leaderboard.
+
+    # Save the evaluation result to the leaderboard
     result = {
         "submission_name": submission_name,
         "model_description": model_description,
@@ -133,7 +120,7 @@ def evaluate_submission(file, submission_name, model_description):
         "mean_weight_error": mean_weight_error,
     }
     save_submission_result(result)
-    
+
     result_text = (
         f"**Evaluation Results for '{submission_name}'**\n\n"
         f"- Model: {model_description}\n"
@@ -148,9 +135,6 @@ LEADERBOARD_FILE = "evaluation_results.csv"
 LEADERBOARD_COLUMNS = ["submission_name", "model_description", "mean_iou", "mean_weight_error"]
 
 def save_submission_result(result):
-    """
-    Save a new submission result into the leaderboard CSV.
-    """
     if os.path.exists(LEADERBOARD_FILE):
         df = pd.read_csv(LEADERBOARD_FILE)
         df = pd.concat([df, pd.DataFrame([result])], ignore_index=True)
@@ -159,9 +143,6 @@ def save_submission_result(result):
     df.to_csv(LEADERBOARD_FILE, index=False)
 
 def load_leaderboard():
-    """
-    Load and sort the leaderboard dataframe.
-    """
     if os.path.exists(LEADERBOARD_FILE):
         df = pd.read_csv(LEADERBOARD_FILE)
         df = df.sort_values(by="mean_weight_error", ascending=True)
@@ -178,27 +159,36 @@ def evaluate_and_refresh(file, submission_name, model_description):
 
 with gr.Blocks() as demo:
     gr.Markdown("# Benchmark Leaderboard: Food Object Detection & Food Weight Prediction")
-    gr.Markdown("This leaderboard evaluates CSV submissions based on bounding box detection (IoU) and food weight predictions (grams).")
-    
+    gr.Markdown(
+        "This leaderboard evaluates CSV submissions based on bounding boxes (IoU) "
+        "and food weight (grams). **Now merging on image_name + class_id** to handle multiple items."
+    )
+
     with gr.Tabs():
         with gr.TabItem("🏅 Leaderboard"):
             leaderboard_output = gr.Dataframe(label="Leaderboard")
             demo.load(load_leaderboard, outputs=leaderboard_output)
             refresh_button = gr.Button("Refresh Leaderboard")
+
             def refresh_leaderboard():
                 return load_leaderboard()
+
             refresh_button.click(refresh_leaderboard, outputs=leaderboard_output)
-        
+
         with gr.TabItem("🚀 Submit CSV"):
-            gr.Markdown("**Submit your predictions CSV file.** The CSV must have the following columns:\n\n"
-                        "`image_id, xmin, ymin, xmax, ymax, weight`")
-            
-            submission_file = gr.File(label="Upload Submission CSV", file_types=[".csv"])
+            gr.Markdown(
+                "**Submit your predictions CSV file.** Required columns:\n\n"
+                "`image_name, class_id, xmin, ymin, xmax, ymax, weight`"
+            )
+
+            submission_file = gr.File(
+                label="Upload Submission CSV", file_types=[".csv"]
+            )
             submission_name_textbox = gr.Textbox(label="Submission Name / Your Name")
             model_description_textbox = gr.Textbox(label="Model Description / Model Name")
             submit_button = gr.Button("Evaluate Submission")
             evaluation_output = gr.Markdown()
-            
+
             submit_button.click(
                 evaluate_and_refresh,
                 inputs=[submission_file, submission_name_textbox, model_description_textbox],