Fix: remove rect boxes from vis, pass per-image crop_offset, only draw lines

app.py

@@ -533,11 +533,11 @@ def _build_complex_visualization(log_entry, patch_info):
         q_orig_size = patch_info.get('query_image_size', (orig_w1, orig_h1))
         h_orig_size = patch_info.get('hist_image_size', (orig_w2, orig_h2))
 
-        # CenterCrop 偏移和 Resize 缩放
-        q_crop = patch_info.get('crop_offset', (0, 0))
-        h_crop = patch_info.get('crop_offset', (0, 0))
-        q_resize = patch_info.get('resize_size', (518, 518))
-        h_resize = patch_info.get('resize_size', (518, 518))
+        # CenterCrop 偏移和 Resize 缩放（query 和 hist 各自独立）
+        q_crop = patch_info.get('query_crop_offset', (0, 0))
+        h_crop = patch_info.get('hist_crop_offset', (0, 0))
+        q_resize = patch_info.get('query_resize_size', (518, 518))
+        h_resize = patch_info.get('hist_resize_size', (518, 518))
 
         q_n_h, q_n_w = q_grid
         h_n_h, h_n_w = h_grid
@@ -553,11 +553,8 @@ def _build_complex_visualization(log_entry, patch_info):
         canvas[:h1, :w1] = img1_disp
         canvas[:h2, w1+gap:] = img2_disp
 
-        def _patch_to_orig_coords(patch_idx, n_w, n_h, crop_offset, scale_resize, display_scale, orig_size, is_right, gap_offset):
-            """将 patch 索引转换为画布上的像素坐标
-
-            步骤：patch_idx → (row,col) → 518x518像素 → 加crop偏移 → 除缩放 → 原图像素 → 乘display_scale → 画布坐标
-            """
+        def _patch_to_center(patch_idx, n_w, crop_offset, scale_resize, display_scale, is_right, gap_offset):
+            """将 patch 索引转换为画布上的中心点像素坐标"""
             row = patch_idx // n_w
             col = patch_idx % n_w
 
@@ -580,44 +577,19 @@ def _build_complex_visualization(log_entry, patch_info):
             if is_right:
                 cx_canvas += gap_offset
 
-            # patch 在原图上的覆盖区域（用于画矩形）
-            x1_518 = col * patch_size + crop_offset[0]
-            y1_518 = row * patch_size + crop_offset[1]
-            x1_orig = x1_518 / scale_resize
-            y1_orig = y1_518 / scale_resize
-            x2_orig = (col + 1) * patch_size + crop_offset[0] / scale_resize
-            y2_orig = (row + 1) * patch_size + crop_offset[1] / scale_resize
-
-            rx1 = int(x1_orig * display_scale)
-            ry1 = int(y1_orig * display_scale)
-            rx2 = int(x2_orig * display_scale)
-            ry2 = int(y2_orig * display_scale)
-
-            if is_right:
-                rx1 += gap_offset
-                rx2 += gap_offset
-
-            return cx_canvas, cy_canvas, rx1, ry1, rx2, ry2
+            return cx_canvas, cy_canvas
 
         for qi, hi, sim in matches:
-            q_cx, q_cy, q_rx1, q_ry1, q_rx2, q_ry2 = \
-                _patch_to_orig_coords(qi, q_n_w, q_n_h, q_crop, q_scale_resize, scale1, q_orig_size, False, 0)
-            h_cx, h_cy, h_rx1, h_ry1, h_rx2, h_ry2 = \
-                _patch_to_orig_coords(hi, h_n_w, h_n_h, h_crop, h_scale_resize, scale2, h_orig_size, True, w1 + gap)
+            q_cx, q_cy = _patch_to_center(qi, q_n_w, q_crop, q_scale_resize, scale1, False, 0)
+            h_cx, h_cy = _patch_to_center(hi, h_n_w, h_crop, h_scale_resize, scale2, True, w1 + gap)
 
             # 颜色：相似度越高越绿，越低越黄
             intensity = min(1.0, max(0.0, (sim - 0.3) / 0.7))
             color = (0, int(200 * intensity + 55), int(255 * (1 - intensity)))
 
-            # 绘制连线
+            # 只绘制连线
             cv2.line(canvas, (q_cx, q_cy), (h_cx, h_cy), color, 1, cv2.LINE_AA)
 
-            # 绘制 patch 矩形（限制在图片范围内）
-            cv2.rectangle(canvas, (max(q_rx1, 0), max(q_ry1, 0)),
-                          (min(q_rx2, w1-1), min(q_ry2, h1-1)), color, 1)
-            cv2.rectangle(canvas, (max(h_rx1, w1+gap), max(h_ry1, 0)),
-                          (min(h_rx2, w1+gap+w2-1), min(h_ry2, h2-1)), color, 1)
-
         n_matches = len(matches)
         avg_sim = sum(s for _, _, s in matches) / max(n_matches, 1)
         cv2.putText(canvas, 'Query', (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0,0,0), 2)

module/config.py

@@ -24,7 +24,7 @@ QDRANT_PATH = os.path.join(PROJECT_ROOT, 'qdrant_data')
 MIN_MATCHES = 20
 MIN_INLIER_RATIO = 0.15
 EDGE_MARGIN = 3
-DINOV2_SIM_THRESHOLD = 0.5
+DINOV2_SIM_THRESHOLD = 0.45
 DINOV2_HIGH_CONF_THRESHOLD = 0.7
 STRICT_MIN_MATCHES = 100
 STRICT_MIN_INLIER_RATIO = 0.50

module/dinov2_utils.py

@@ -205,4 +205,8 @@ class DINOv2Extractor:
             'hist_grid': feat2['patch_grid'],
             'query_image_size': feat1['image_size'],
             'hist_image_size': feat2['image_size'],
+            'query_crop_offset': feat1.get('crop_offset', (0, 0)),
+            'hist_crop_offset': feat2.get('crop_offset', (0, 0)),
+            'query_resize_size': feat1.get('resize_size', (518, 518)),
+            'hist_resize_size': feat2.get('resize_size', (518, 518)),
         }