scorevision: push artifact

miner.py

@@ -479,16 +479,29 @@ class Miner:
             ))
         return out
 
-    # ── Person preprocessing (stretch resize) ──────────────────────────────
+    # ── Person preprocessing (letterbox) ──────────────────────────────────
+
+    def _per_letterbox(self, img):
+        h, w = img.shape[:2]
+        r = min(self.per_h / h, self.per_w / w)
+        nw, nh = int(round(w * r)), int(round(h * r))
+        img_r = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_LINEAR)
+        dw, dh = self.per_w - nw, self.per_h - nh
+        pl, pt = dw // 2, dh // 2
+        img_p = cv2.copyMakeBorder(
+            img_r, pt, dh - pt, pl, dw - pl,
+            cv2.BORDER_CONSTANT, value=(114, 114, 114),
+        )
+        return img_p, r, pl, pt
 
     def _per_preprocess(self, image_bgr):
-        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-        resized = cv2.resize(rgb, (self.per_w, self.per_h))
-        x = resized.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))[None, ...]
-        return x
+        img_p, ratio, pl, pt = self._per_letterbox(image_bgr)
+        rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
+        inp = rgb.astype(np.float32) / 255.0
+        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
+        return inp, ratio, pl, pt
 
-    def _per_decode(self, raw, oh, ow, conf_thresh):
+    def _per_decode(self, raw, ratio, pl, pt, oh, ow, conf_thresh):
         pred = raw[0]
         if pred.ndim != 2:
             return np.empty((0, 4)), np.empty(0)
@@ -501,9 +514,10 @@ class Miner:
             boxes, confs = pred[keep, :4], confs[keep]
             if len(boxes) == 0:
                 return np.empty((0, 4)), np.empty(0)
-            sx, sy = ow / float(self.per_w), oh / float(self.per_h)
-            boxes[:, 0] *= sx; boxes[:, 2] *= sx
-            boxes[:, 1] *= sy; boxes[:, 3] *= sy
+            boxes[:, 0] = (boxes[:, 0] - pl) / ratio
+            boxes[:, 2] = (boxes[:, 2] - pl) / ratio
+            boxes[:, 1] = (boxes[:, 1] - pt) / ratio
+            boxes[:, 3] = (boxes[:, 3] - pt) / ratio
             boxes = np.clip(boxes, 0, [[ow, oh, ow, oh]])
             return boxes, confs
 
@@ -518,19 +532,18 @@ class Miner:
         boxes, confs = pred[keep, :4], confs[keep]
         if len(boxes) == 0:
             return np.empty((0, 4)), np.empty(0)
-        sx, sy = ow / float(self.per_w), oh / float(self.per_h)
         cx, cy, bw, bh = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
-        x1 = np.clip((cx - bw / 2) * sx, 0, ow)
-        y1 = np.clip((cy - bh / 2) * sy, 0, oh)
-        x2 = np.clip((cx + bw / 2) * sx, 0, ow)
-        y2 = np.clip((cy + bh / 2) * sy, 0, oh)
+        x1 = np.clip((cx - bw / 2 - pl) / ratio, 0, ow)
+        y1 = np.clip((cy - bh / 2 - pt) / ratio, 0, oh)
+        x2 = np.clip((cx + bw / 2 - pl) / ratio, 0, ow)
+        y2 = np.clip((cy + bh / 2 - pt) / ratio, 0, oh)
         return np.stack([x1, y1, x2, y2], axis=1), confs
 
     def _per_run_pass(self, image_bgr, conf_thresh):
         oh, ow = image_bgr.shape[:2]
-        inp = self._per_preprocess(image_bgr)
+        inp, ratio, pl, pt = self._per_preprocess(image_bgr)
         raw = self.per_session.run(None, {self.per_input_name: inp})[0]
-        return self._per_decode(raw, oh, ow, conf_thresh)
+        return self._per_decode(raw, ratio, pl, pt, oh, ow, conf_thresh)
 
     def _infer_person(self, image_bgr):
         oh, ow = image_bgr.shape[:2]
@@ -713,4 +726,4 @@ class Miner:
         ).start()
 
         return results
-# Miner v3.1 — repo-split + YOLO26s person 960px 20260331
+# Miner v3.2 — letterbox person preprocessing + YOLO26s 960px 20260331