marginalized blobs with marginalized tiles

inference_tab/helpers.py

@@ -235,15 +235,12 @@ def processYOLOBoxes(iou):
     boxes_full = [b["bbox"] for b in box_data]  # Format: [x1, y1, x2, y2]
     return boxes_full
 
-def prepare_tiles(image_path, boxes_full, tile_size=1024, overlap=50, iou=0.5, c_th=0.75, edge_margin=10, tile_margin=300):
+def prepare_tiles(image_path, boxes_full, tile_size=1024, overlap=256, iou=0.5, c_th=0.75, edge_margin=10):
     """
     Tiles the image and prepares per-tile metadata including filtered boxes and point prompts.
     Returns full image size H, W.
     """
 
-    full_img = cv2.imread(image_path)
-    H_full, W_full, _ = full_img.shape
-
     tiles, (H, W, _) = tile_image_with_overlap(image_path, tile_size, overlap)
     os.makedirs("tmp/tiles", exist_ok=True)
     meta = []
@@ -253,14 +250,6 @@ def prepare_tiles(image_path, boxes_full, tile_size=1024, overlap=50, iou=0.5, c
         tile_array = cv2.cvtColor(tile_array, cv2.COLOR_BGR2RGB)
         Image.fromarray(tile_array).save(tile_path)
 
-        my1 = max(0, y_offset - tile_margin)
-        my2 = min(H_full, y_offset + tile_size + tile_margin)
-        mx1 = max(0, x_offset - tile_margin)
-        mx2 = min(W_full, x_offset + tile_size + tile_margin)
-        tile_with_margin = full_img[my1:my2, mx1:mx2]
-        cv2.imwrite(f"tiles/tile_{idx}_margin.png", tile_with_margin)
-
-
         tile_h, tile_w, _ = tile_array.shape
 
         # Select boxes overlapping this tile
@@ -342,7 +331,6 @@ def prepare_tiles(image_path, boxes_full, tile_size=1024, overlap=50, iou=0.5, c
 
 
 
-
 def merge_tile_masks(H, W):
     """
     Merge predicted tile masks into a full-size image.

inference_tab/inference_logic.py

@@ -37,11 +37,12 @@ _trocr_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 
 
+
 def run_inference(tile_dict, gcp_path,user_crs, city_name, score_th, hist_th, hist_dic):
     IMAGE_FOLDER = os.path.join(OUTPUT_DIR, "blobs")
     CSV_FILE = os.path.join(OUTPUT_DIR, "annotations.csv")
     MASK_FILE = os.path.join(OUTPUT_DIR, "mask.tif")
-    
+
 
     if os.path.exists(IMAGE_FOLDER):
         shutil.rmtree(IMAGE_FOLDER)
@@ -89,7 +90,7 @@ def run_inference(tile_dict, gcp_path,user_crs, city_name, score_th, hist_th, hi
     if stop_pipeline:
         yield log + "Pipeline stopped: no text segments found.\n", None
         return
-
+    
     for msg in extractSegments(image_path):
         log += msg + "\n"
         yield log, None
@@ -275,7 +276,7 @@ def getSegments(image_path,iou=0.5,c_th=0.75,edge_margin=10):
     with open(BOXES_PATH, "r") as f:
         box_data = json.load(f)
     boxes = [b["bbox"] for b in box_data]
-    yield "Prepare tiles..."
+    yield f"Prepare tiles for {image_path}..."
     H,W = prepare_tiles(image_path, boxes, tile_size=1024, overlap=50, iou=iou, c_th=c_th, edge_margin=edge_margin)
     yield "Run inference on tiles..."
     for msg in run_tile_inference():
@@ -283,23 +284,16 @@ def getSegments(image_path,iou=0.5,c_th=0.75,edge_margin=10):
         if "Stopping inference" in msg:
             yield "No labels detected – halting getSegments."
             return
-    yield "Marge predicted masks into image..."
+    yield "Merge predicted masks into image..."
     merge_tile_masks(H,W)
 
     MASK_PATH = os.path.join(OUTPUT_DIR,"mask.tif")
     yield f"{MASK_PATH}"
-    
 
 
 def extractSegments(image_path, min_size=500, margin=150):
     image = cv2.imread(image_path)
     
-    margin_path = image_path.replace(".png", "_margin.png")
-    image_margin_tile = cv2.imread(margin_path)
-    if image_margin_tile is None:
-        yield f"Error: Could not load margin tile at {margin_path}"
-
-
     MASK_PATH = os.path.join(OUTPUT_DIR, "mask.tif")
     mask = cv2.imread(MASK_PATH, cv2.IMREAD_UNCHANGED)
 
@@ -337,22 +331,47 @@ def extractSegments(image_path, min_size=500, margin=150):
         y_max_m = min(height, y_max + margin)
 
         cropped_image_margin = image[y_min_m:y_max_m, x_min_m:x_max_m]
-        cropped_mask_margin = blob_mask[y_min_m:y_max_m, x_min_m:x_max_m]'''
-
-        cropped_image_margin = image_margin_tile[y_min : y_max + 2*margin, 
-                                                x_min : x_max + 2*margin]
-        cropped_mask_margin = np.zeros(cropped_image_margin.shape[:2], dtype=np.uint8)
-        cropped_mask_margin[margin : margin + (y_max-y_min), 
-                           margin : margin + (x_max-x_min)] = blob_mask[y_min:y_max, x_min:x_max]
+        cropped_mask_margin = blob_mask[y_min_m:y_max_m, x_min_m:x_max_m]
 
         shaded_margin = cropped_image_margin.copy()
         overlay_margin = cropped_image_margin.copy()
         overlay_margin[cropped_mask_margin == 1] = (255, 200, 100)
         shaded_margin = cv2.addWeighted(overlay_margin, 0.35, shaded_margin, 0.65, 0)
 
+        BLOB_PATH_MARGIN = os.path.join(OUTPUT_DIR, "blobs", f"{blob_id}_margin.png")
+        cv2.imwrite(BLOB_PATH_MARGIN, shaded_margin)'''
+
+        # ---- MARGINALIZED NEW (with mask/shading) ----
+        base, ext = os.path.splitext(image_path)
+        margin_tile_path = f"{base}_margin{ext}"
+        margin_img = cv2.imread(margin_tile_path)
+        h_m, w_m = margin_img.shape[:2]
+
+        x1_m = max(0, x_min)
+        y1_m = max(0, y_min)
+        x2_m = min(w_m, x_max + (2 * margin))
+        y2_m = min(h_m, y_max + (2 * margin))
+        cropped = margin_img[y1_m:y2_m, x1_m:x2_m]
+
+        padded_blob_mask = cv2.copyMakeBorder(
+            blob_mask, 
+            margin, margin, margin, margin, 
+            borderType=cv2.BORDER_CONSTANT, 
+            value=0
+        )
+
+        cropped_mask = padded_blob_mask[y1_m:y2_m, x1_m:x2_m]
+
+        shaded_margin = cropped.copy()
+        overlay_margin = cropped.copy()
+        
+        overlay_margin[cropped_mask == 1] = (255, 200, 100)
+        shaded_margin = cv2.addWeighted(overlay_margin, 0.35, shaded_margin, 0.65, 0)
+
         BLOB_PATH_MARGIN = os.path.join(OUTPUT_DIR, "blobs", f"{blob_id}_margin.png")
         cv2.imwrite(BLOB_PATH_MARGIN, shaded_margin)
 
+
     yield f"Done."
 
 

inference_tab/inference_setup.py

@@ -5,6 +5,7 @@ from PIL import Image
 import os
 
 TILE_SIZE = 1024  
+MARGIN_TILE = 150
 TILE_FOLDER = "tiles"
 os.makedirs(TILE_FOLDER, exist_ok=True)
 tiles_cache = {"tiles": [], "selected_tile": None}
@@ -29,13 +30,27 @@ def make_tiles(image, tile_size=TILE_SIZE):
 def create_tiles(image_file):
     img = Image.open(image_file.name).convert("RGB")
     img = np.array(img)
+    H, W, _ = img.shape
 
     annotated, tiles = make_tiles(img, TILE_SIZE)
     tiles_cache["tiles"] = []
 
     for idx, (coords, tile) in enumerate(tiles):
+        x, y, x_end, y_end = coords
+
         tile_path = os.path.join(TILE_FOLDER, f"tile_{idx}.png")
         Image.fromarray(tile).save(tile_path)
+
+        # Marginalized tile
+        x_m = max(0, x - MARGIN_TILE)
+        y_m = max(0, y - MARGIN_TILE)
+        x_end_m = min(W, x_end + MARGIN_TILE)
+        y_end_m = min(H, y_end + MARGIN_TILE)
+        tile_margin = img[y_m:y_end_m, x_m:x_end_m]
+        tile_margin_path = os.path.join(TILE_FOLDER, f"tile_{idx}_margin.png")
+        Image.fromarray(tile_margin).save(tile_margin_path)
+
+
         tiles_cache["tiles"].append((coords, tile_path))  # store path instead of array
 
     tiles_cache["selected_tile"] = None