fixing NMSBoxesBatched function accepts Rect objects instead of two diagonal corner coordinates (#296)

* fix: fixing NMSBoxesBatched function accepts Rect objects instead of two diagonal corner coordinates

* feat: move to hf (#297)

* Revert "fix: fixing NMSBoxesBatched function accepts Rect objects instead of two diagonal corner coordinates"

This reverts commit d708146.

	modified:   models/object_detection_yolox/demo.cpp
modified:   models/object_detection_yolox/demo.py
	modified:   models/object_detection_yolox/yolox.py

* revert: modify to best conform to the original code style

---------

Co-authored-by: cuixing158 <1229261804@qq.com>
Co-authored-by: Yuantao Feng <yuantao.feng@opencv.org.cn>

Author: cuixing158

models/object_detection_yolox/demo.cpp

@@ -104,38 +104,38 @@ class YoloX {
             boxes.rowRange(r, r + 1) = *ptr * boxes.rowRange(r, r + 1);
         }
         // get boxes
-        Mat boxes_xyxy(boxes.rows, boxes.cols, CV_32FC1, Scalar(1));
+        Mat boxes_xywh(boxes.rows, boxes.cols, CV_32FC1, Scalar(1));
         Mat scores = dets.colRange(5, dets.cols).clone();
         vector<float> maxScores(dets.rows);
         vector<int> maxScoreIdx(dets.rows);
-        vector<Rect2d> boxesXYXY(dets.rows);
+        vector<Rect2d> boxesXYWH(dets.rows);
 
-        for (int r = 0; r < boxes_xyxy.rows; r++, ptr++)
+        for (int r = 0; r < boxes_xywh.rows; r++, ptr++)
         {
-            boxes_xyxy.at<float>(r, 0) = boxes.at<float>(r, 0) - boxes.at<float>(r, 2) / 2.f;
-            boxes_xyxy.at<float>(r, 1) = boxes.at<float>(r, 1) - boxes.at<float>(r, 3) / 2.f;
-            boxes_xyxy.at<float>(r, 2) = boxes.at<float>(r, 0) + boxes.at<float>(r, 2) / 2.f;
-            boxes_xyxy.at<float>(r, 3) = boxes.at<float>(r, 1) + boxes.at<float>(r, 3) / 2.f;
+            boxes_xywh.at<float>(r, 0) = boxes.at<float>(r, 0) - boxes.at<float>(r, 2) / 2.f;
+            boxes_xywh.at<float>(r, 1) = boxes.at<float>(r, 1) - boxes.at<float>(r, 3) / 2.f;
+            boxes_xywh.at<float>(r, 2) = boxes.at<float>(r, 2);
+            boxes_xywh.at<float>(r, 3) = boxes.at<float>(r, 3);
             // get scores and class indices
             scores.rowRange(r, r + 1) = scores.rowRange(r, r + 1) * dets.at<float>(r, 4);
             double minVal, maxVal;
             Point maxIdx;
             minMaxLoc(scores.rowRange(r, r+1), &minVal, &maxVal, nullptr, &maxIdx);
             maxScoreIdx[r] = maxIdx.x;
             maxScores[r] = float(maxVal);
-            boxesXYXY[r].x = boxes_xyxy.at<float>(r, 0);
-            boxesXYXY[r].y = boxes_xyxy.at<float>(r, 1);
-            boxesXYXY[r].width = boxes_xyxy.at<float>(r, 2);
-            boxesXYXY[r].height = boxes_xyxy.at<float>(r, 3);
+            boxesXYWH[r].x = boxes_xywh.at<float>(r, 0);
+            boxesXYWH[r].y = boxes_xywh.at<float>(r, 1);
+            boxesXYWH[r].width = boxes_xywh.at<float>(r, 2);
+            boxesXYWH[r].height = boxes_xywh.at<float>(r, 3);
         }
 
         vector<int> keep;
-        NMSBoxesBatched(boxesXYXY, maxScores, maxScoreIdx, this->confThreshold, this->nmsThreshold, keep);
+        NMSBoxesBatched(boxesXYWH, maxScores, maxScoreIdx, this->confThreshold, this->nmsThreshold, keep);
         Mat candidates(int(keep.size()), 6, CV_32FC1);
         int row = 0;
         for (auto idx : keep)
         {
-            boxes_xyxy.rowRange(idx, idx + 1).copyTo(candidates(Rect(0, row, 4, 1)));
+            boxes_xywh.rowRange(idx, idx + 1).copyTo(candidates(Rect(0, row, 4, 1)));
             candidates.at<float>(row, 4) = maxScores[idx];
             candidates.at<float>(row, 5) = float(maxScoreIdx[idx]);
             row++;

models/object_detection_yolox/demo.py

@@ -58,12 +58,12 @@ def vis(dets, srcimg, letterbox_scale, fps=None):
         score = det[-2]
         cls_id = int(det[-1])
 
-        x0, y0, x1, y1 = box
+        x0, y0, w, h = box
 
         text = '{}:{:.1f}%'.format(classes[cls_id], score * 100)
         font = cv.FONT_HERSHEY_SIMPLEX
         txt_size = cv.getTextSize(text, font, 0.4, 1)[0]
-        cv.rectangle(res_img, (x0, y0), (x1, y1), (0, 255, 0), 2)
+        cv.rectangle(res_img, (x0, y0 , w, h), (0, 255, 0), 2)
         cv.rectangle(res_img, (x0, y0 + 1), (x0 + txt_size[0] + 1, y0 + int(1.5 * txt_size[1])), (255, 255, 255), -1)
         cv.putText(res_img, text, (x0, y0 + txt_size[1]), font, 0.4, (0, 0, 0), thickness=1)
 

models/object_detection_yolox/yolox.py

@@ -50,20 +50,20 @@ def postprocess(self, outputs):
 
         # get boxes
         boxes = dets[:, :4]
-        boxes_xyxy = np.ones_like(boxes)
-        boxes_xyxy[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.
-        boxes_xyxy[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.
-        boxes_xyxy[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.
-        boxes_xyxy[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.
+        boxes_xywh = np.ones_like(boxes)
+        boxes_xywh[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.
+        boxes_xywh[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.
+        boxes_xywh[:, 2] = boxes[:, 2]
+        boxes_xywh[:, 3] = boxes[:, 3]
 
         # get scores and class indices
         scores = dets[:, 4:5] * dets[:, 5:]
         max_scores = np.amax(scores, axis=1)
         max_scores_idx = np.argmax(scores, axis=1)
 
-        keep = cv2.dnn.NMSBoxesBatched(boxes_xyxy.tolist(), max_scores.tolist(), max_scores_idx.tolist(), self.confThreshold, self.nmsThreshold)
+        keep = cv2.dnn.NMSBoxesBatched(boxes_xywh.tolist(), max_scores.tolist(), max_scores_idx.tolist(), self.confThreshold, self.nmsThreshold)
 
-        candidates = np.concatenate([boxes_xyxy, max_scores[:, None], max_scores_idx[:, None]], axis=1)
+        candidates = np.concatenate([boxes_xywh, max_scores[:, None], max_scores_idx[:, None]], axis=1)
         if len(keep) == 0:
             return np.array([])
         return candidates[keep]