From 8dcbfb5535826a264403dffc524b812ca2924f40 Mon Sep 17 00:00:00 2001
From: Admin <admin@viaeon.com>
Date: Fri, 17 Jul 2026 08:54:23 +0000
Subject: [PATCH] fix: handle YOLO11 output format [1,5,8400] with transpose
 for OpenCV DNN

---
 pkg/opencv/opencv.cpp | 55 +++++++++++++++++++++++++++++++++----------
 1 file changed, 43 insertions(+), 12 deletions(-)

diff --git a/pkg/opencv/opencv.cpp b/pkg/opencv/opencv.cpp
index 873f065..2ed5800 100644
--- a/pkg/opencv/opencv.cpp
+++ b/pkg/opencv/opencv.cpp
@@ -165,24 +165,55 @@ int cv_yolo_detect(const char* name, const unsigned char* img_data, int width, i
         float scaleY = (float)frame.rows / inpHeight;
         
         for (size_t i = 0; i < outs.size(); ++i) {
-            float* data = (float*)outs[i].data;
-            for (int j = 0; j < outs[i].rows; ++j, data += outs[i].cols) {
-                Mat scores = outs[i].row(j).colRange(5, outs[i].cols);
-                Point classIdPoint;
-                double confidence;
-                minMaxLoc(scores, 0, &confidence, 0, &classIdPoint);
+            // YOLO11 输出格式: [1, 5+num_classes, 8400] (通道在前)
+            // 需要转置为 [8400, 5+num_classes]
+            Mat out = outs[i];
+            int num_detections = out.size[2];  // 8400
+            int num_features = out.size[1];     // 5 + num_classes
+            
+            // 转置输出
+            Mat out_transposed;
+            if (out.size[1] < out.size[2]) {
+                // [1, features, detections] -> [detections, features]
+                Mat out_2d(out.size[1], out.size[2], CV_32F, out.data);
+                transpose(out_2d, out_transposed);
+            } else {
+                out_transposed = out;
+            }
+            
+            for (int j = 0; j < num_detections; ++j) {
+                float* data = out_transposed.ptr<float>(j);
+                
+                float obj_conf = data[4];
+                if (obj_conf < detector->confThreshold) continue;
+                
+                // 找最大类别分数
+                int classId = 0;
+                float maxScore = 0;
+                for (int k = 5; k < num_features; ++k) {
+                    if (data[k] > maxScore) {
+                        maxScore = data[k];
+                        classId = k - 5;
+                    }
+                }
+                float confidence = obj_conf * maxScore;
                 
                 if (confidence > detector->confThreshold) {
-                    int centerX = (int)(data[0] * scaleX);
-                    int centerY = (int)(data[1] * scaleY);
-                    int width = (int)(data[2] * scaleX);
-                    int height = (int)(data[3] * scaleY);
+                    float cx = data[0];
+                    float cy = data[1];
+                    float w = data[2];
+                    float h = data[3];
+                    
+                    int centerX = (int)(cx * scaleX);
+                    int centerY = (int)(cy * scaleY);
+                    int width = (int)(w * scaleX);
+                    int height = (int)(h * scaleY);
                     
                     int left = centerX - width / 2;
                     int top = centerY - height / 2;
                     
-                    classIds.push_back(classIdPoint.x);
-                    confidences.push_back((float)confidence);
+                    classIds.push_back(classId);
+                    confidences.push_back(confidence);
                     boxes.push_back(Rect(left, top, width, height));
                 }
             }