Working state.

annotating_with_checks.py

@@ -87,7 +87,6 @@ def process_label(root_dir, student_id, label, content):
 
     if not os.path.exists(pdf_path):
         return None, []
-
     base_img, total_h, max_w = annotating.make_base_image(pdf_path)
     if not base_img:
         return None, []
@@ -239,12 +238,14 @@ def process_student(args):
     # Au moins, le drift n'empire pas au fil de la copie
     temp_img_path = os.path.join(output_dir, "Reference.jpg") # Can't use png here
     concat_img.save(temp_img_path, quality=90)
+
     pdf_path = os.path.join(output_dir, "Concat.pdf")
     w, h = concat_img.size
     c = canvas.Canvas(pdf_path, pagesize=(w, h))
     c.drawImage(temp_img_path, 0, 0, width=w, height=h)
     c.save()
 
+    print("Debug : size", w, h)
     # Ancien code, avec du drift
     # concat_img.save(os.path.join(output_dir, "Concat.pdf"), "PDF", resolution=DPI)
     # concat_img.save(os.path.join(output_dir, "Reference.png"))
@@ -268,4 +269,10 @@ if __name__ == "__main__":
     # print(tasks)
 
     with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
-        executor.map(process_student, tasks)
+        results = executor.map(process_student, tasks)
+        try:
+            for _ in results:
+                pass
+        except Exception:
+            import traceback
+            traceback.print_exc()

correction.py

@@ -199,14 +199,6 @@ def generate_request(file, full_label):
         max_output_tokens=65535,
         response_mime_type= "application/json",
         response_json_schema= TypeAdapter(List[EvaluationEntry]).json_schema()
-        # Thinking config is not compatible with response_json ? Unsure.
-        # thinking_config=types.ThinkingConfig(
-          # thinking_budget=-1,
-        # ),
-        # thinking_config=types.ThinkingConfig(
-            # include_thoughts=True,
-            # thinking_budget=1024, # Optimized for Gemini 3 capabilities
-        # ),
     )
     return (contents, generate_content_config)
 

reading_annotations.py

@@ -3,7 +3,7 @@ import os
 import json
 import numpy as np
 import shutil
-from PIL import Image, ImageChops
+from PIL import Image, ImageChops, ImageFilter
 Image.MAX_IMAGE_PIXELS = None
 from pdf2image import convert_from_path
 import annotating # Reuse rendering logic
@@ -36,7 +36,10 @@ def detect_checks_and_notes(output_dir):
     # Warning: If the PDF is huge, pdf2image might split pages or OOM.
     # Assuming user didn't change page dimensions/order.
     try:
-        user_pages = convert_from_path(pdf_path, dpi=DPI)
+        # user_pages = convert_from_path(pdf_path, dpi=DPI)
+        # La version suivante évite les size mismatch
+        # Mais donne plus de bruit
+        user_pages = convert_from_path(pdf_path, dpi=72)
     except Exception as e:
         print(f"Error reading PDF: {e}")
         return [], None
@@ -97,19 +100,21 @@ def detect_checks_and_notes(output_dir):
             # Expand mask slightly to catch sloppy ticks
             mask_draw.rectangle([x1-5, y1-5, x2+5, y2+5], fill=0)
         else:
-            # print("A box, not checked !", density)
+            mask_draw.rectangle([x1-2, y1-2, x2+2, y2+2], fill=0)
-            # Even if not "checked", mask the box area slightly to avoid
+
-            # artifacts if user hovered over it, though arguably we keep it.
+        if box["type"] == "score" and box["value"] == 0.0:
-            # Let's strictly mask only if checked to verify detection?
+            # Mask the whole line
-            # No, prompt says "not extract the part that are just checking".
+            mask_draw.rectangle([0, y1-5, ref_img.width, y2+5], fill=0)
-            # If user checked it, we mask it.
-            pass
 
     # --- Extraction Phase ---
     # Create the "Manual Notes" layer
     # Logic: User - Ref. If Diff is dark -> Note.
     # We want a transparent image with just the pen strokes.
 
+    # Try Gaussian Blur, peut-être inutile.
+    ref_blur = ref_img.filter(ImageFilter.GaussianBlur(5))
+    user_blur = user_img.filter(ImageFilter.GaussianBlur(5))
+
     # 1. Get difference image
     diff_img = ImageChops.difference(ref_img, user_img).convert("L")
 
@@ -117,7 +122,8 @@ def detect_checks_and_notes(output_dir):
     # Pixels that are different enough:
     diff_data = np.array(diff_img)
     # Create alpha channel: 0 where no diff, 255 where diff
-    alpha = np.where(diff_data > 20, 255, 0).astype(np.uint8)
+    # Higher treshold is better
+    alpha = np.where(diff_data > 100, 255, 0).astype(np.uint8)
 
     # 3. Create output image (Black strokes, variable alpha)
     # Or Copy user colors? Better to copy user pixels.