Blurサイズ問題の修正

server/main.py

@@ -83,7 +83,6 @@ class GenerateRequest(BaseModel):
     end_frame: int
     conf_threshold: float = 0.5
     iou_threshold: float = 0.45
-    mask_scale: float = 1.5
 
 
 class BakeRequest(BaseModel):
@@ -91,6 +90,7 @@ class BakeRequest(BaseModel):
     detections_path: str
     output_path: str
     blur_size: int = 50
+    display_scale: float = 1.0
     format: str = "mp4"
 
 
@@ -305,20 +305,15 @@ def process_video_task(task_id: str, req: GenerateRequest):
             for detections in batch_detections:
                 packed_detections: List[List[float]] = []
                 for x, y, w, h, conf in detections:
-                    scaled = _scale_bbox(
-                        int(x),
-                        int(y),
-                        int(w),
-                        int(h),
-                        float(req.mask_scale),
-                        width,
-                        height,
-                    )
-                    if scaled is None:
+                    # bboxをそのまま保存（表示スケールはBake時に適用）
+                    bx, by, bw, bh = int(x), int(y), int(w), int(h)
+                    bx = max(0, bx)
+                    by = max(0, by)
+                    bw = min(width - bx, bw)
+                    bh = min(height - by, bh)
+                    if bw <= 0 or bh <= 0:
                         continue
-                    packed_detections.append(
-                        [scaled[0], scaled[1], scaled[2], scaled[3], float(conf)]
-                    )
+                    packed_detections.append([bx, by, bw, bh, float(conf)])
                 frame_detections.append(packed_detections)
                 current_count += 1
                 tasks[task_id].progress = current_count
@@ -356,7 +351,7 @@ def process_video_task(task_id: str, req: GenerateRequest):
                 "width": width,
                 "height": height,
                 "fps": fps,
-                "mask_scale": float(req.mask_scale),
+                "mask_scale": 1.0,
                 "frames": frame_detections,
             }
             with open(output_msgpack_path, "wb") as f:
@@ -435,9 +430,9 @@ def process_bake_task(task_id: str, req: BakeRequest):
         blur_size = max(1, int(req.blur_size))
         if blur_size % 2 == 0:
             blur_size += 1
-        feather_radius = max(3, min(25, blur_size // 3))
-        feather_kernel = feather_radius * 2 + 1
-        blur_margin = max(1, (blur_size // 2) + feather_radius)
+        display_scale = max(0.1, float(req.display_scale))
+        # blur_margin は境界問題回避のための計算用余白のみ（表示には使わない）
+        blur_margin = blur_size // 2
 
         # Queues
         queue_size = 8
@@ -507,29 +502,37 @@ def process_bake_task(task_id: str, req: BakeRequest):
                         continue
 
                     for x, y, w, h in valid_boxes:
-                        roi_x1 = max(0, x - blur_margin)
-                        roi_y1 = max(0, y - blur_margin)
-                        roi_x2 = min(src_width, x + w + blur_margin)
-                        roi_y2 = min(src_height, y + h + blur_margin)
+                        # display_scale で表示サイズを決定
+                        cx = x + w / 2
+                        cy = y + h / 2
+                        dw = max(1, int(w * display_scale))
+                        dh = max(1, int(h * display_scale))
+                        dx = int(cx - dw / 2)
+                        dy = int(cy - dh / 2)
+
+                        # ROIは表示サイズ + blur_margin（計算用余白、境界問題回避のみ）
+                        roi_x1 = max(0, dx - blur_margin)
+                        roi_y1 = max(0, dy - blur_margin)
+                        roi_x2 = min(src_width, dx + dw + blur_margin)
+                        roi_y2 = min(src_height, dy + dh + blur_margin)
                         roi_width = roi_x2 - roi_x1
                         roi_height = roi_y2 - roi_y1
 
                         if roi_width <= 0 or roi_height <= 0:
                             continue
 
-                        roi_mask = np.zeros((roi_height, roi_width), dtype=np.uint8)
-                        center = (x + w // 2 - roi_x1, y + h // 2 - roi_y1)
-                        axes = (max(1, w // 2), max(1, h // 2))
-                        cv2.ellipse(roi_mask, center, axes, 0, 0, 360, 255, -1)
-
-                        roi_mask = cv2.GaussianBlur(roi_mask, (feather_kernel, feather_kernel), 0)
+                        # ブラーはROI全体で計算（余白があるので端の精度が保証される）
                         roi_src = frame[roi_y1:roi_y2, roi_x1:roi_x2]
                         roi_blurred = cv2.GaussianBlur(roi_src, (blur_size, blur_size), 0)
 
+                        # 合成マスクはdisplay_scaleサイズの楕円のみ（featheringなし）
+                        roi_mask = np.zeros((roi_height, roi_width), dtype=np.uint8)
+                        center = (int(cx) - roi_x1, int(cy) - roi_y1)
+                        axes = (max(1, dw // 2), max(1, dh // 2))
+                        cv2.ellipse(roi_mask, center, axes, 0, 0, 360, 255, -1)
+
                         roi_alpha = (roi_mask.astype(np.float32) / 255.0)[..., np.newaxis]
-                        roi_composed = (roi_src.astype(np.float32) * (1.0 - roi_alpha)) + (
-                            roi_blurred.astype(np.float32) * roi_alpha
-                        )
+                        roi_composed = roi_src.astype(np.float32) * (1.0 - roi_alpha) + roi_blurred.astype(np.float32) * roi_alpha
                         frame[roi_y1:roi_y2, roi_x1:roi_x2] = np.clip(roi_composed, 0, 255).astype(np.uint8)
 
                     process_queue.put((idx, frame))