update retinaface_trt del cvtColor and increase THRESH (wang-xinyu#535)

* update retinaface_trt del cvtColor and increase THRESH

* update macro in retinafce_xx.cpp and preprocess_image description

* retinaface_xx.cpp add a space after

* HRESH for CONF and THRESHOLD for IOU is unified --THRESH

Co-authored-by: Admin <[email protected]>

Author: Huang ZhuangZhuang

retinaface/retina_mnet.cpp

@@ -12,7 +12,8 @@
 #define USE_FP16  // set USE_INT8 or USE_FP16 or USE_FP32
 #define DEVICE 0  // GPU id
 #define BATCH_SIZE 1
-#define VIS_THRESH 0.6
+#define CONF_THRESH = 0.75
+#define IOU_THRESH = 0.4
 
 // stuff we know about the network and the input/output blobs
 static const int INPUT_H = decodeplugin::INPUT_H;  // H, W must be able to  be divided by 32.
@@ -349,13 +350,13 @@ int main(int argc, char** argv) {
 
     for (int b = 0; b < BATCH_SIZE; b++) {
         std::vector<decodeplugin::Detection> res;
-        nms(res, &prob[b * OUTPUT_SIZE]);
+        nms(res, &prob[b * OUTPUT_SIZE], IOU_THRESH);
         std::cout << "number of detections -> " << prob[b * OUTPUT_SIZE] << std::endl;
         std::cout << " -> " << prob[b * OUTPUT_SIZE + 10] << std::endl;
         std::cout << "after nms -> " << res.size() << std::endl;
         cv::Mat tmp = img.clone();
         for (size_t j = 0; j < res.size(); j++) {
-            if (res[j].class_confidence < VIS_THRESH) continue;
+            if (res[j].class_confidence < CONF_THRESH) continue;
             cv::Rect r = get_rect_adapt_landmark(tmp, INPUT_W, INPUT_H, res[j].bbox, res[j].landmark);
             cv::rectangle(tmp, r, cv::Scalar(0x27, 0xC1, 0x36), 2);
             //cv::putText(tmp, std::to_string((int)(res[j].class_confidence * 100)) + "%", cv::Point(r.x, r.y - 1), cv::FONT_HERSHEY_PLAIN, 1.2, cv::Scalar(0xFF, 0xFF, 0xFF), 1);

retinaface/retina_r50.cpp

@@ -12,6 +12,8 @@
 #define USE_INT8  // set USE_INT8 or USE_FP16 or USE_FP32
 #define DEVICE 0  // GPU id
 #define BATCH_SIZE 1
+#define CONF_THRESH = 0.75
+#define IOU_THRESH = 0.4
 
 // stuff we know about the network and the input/output blobs
 static const int INPUT_H = decodeplugin::INPUT_H;  // H, W must be able to  be divided by 32.
@@ -369,13 +371,13 @@ int main(int argc, char** argv) {
 
     for (int b = 0; b < BATCH_SIZE; b++) {
         std::vector<decodeplugin::Detection> res;
-        nms(res, &prob[b * OUTPUT_SIZE]);
+        nms(res, &prob[b * OUTPUT_SIZE], IOU_THRESH);
         std::cout << "number of detections -> " << prob[b * OUTPUT_SIZE] << std::endl;
         std::cout << " -> " << prob[b * OUTPUT_SIZE + 10] << std::endl;
         std::cout << "after nms -> " << res.size() << std::endl;
         cv::Mat tmp = img.clone();
         for (size_t j = 0; j < res.size(); j++) {
-            if (res[j].class_confidence < 0.1) continue;
+            if (res[j].class_confidence < CONF_THRESH) continue;
             cv::Rect r = get_rect_adapt_landmark(tmp, INPUT_W, INPUT_H, res[j].bbox, res[j].landmark);
             cv::rectangle(tmp, r, cv::Scalar(0x27, 0xC1, 0x36), 2);
             //cv::putText(tmp, std::to_string((int)(res[j].class_confidence * 100)) + "%", cv::Point(r.x, r.y - 1), cv::FONT_HERSHEY_PLAIN, 1.2, cv::Scalar(0xFF, 0xFF, 0xFF), 1);

retinaface/retinaface_trt.py

@@ -19,8 +19,8 @@
 
 INPUT_H = 480  #defined in decode.h
 INPUT_W = 640
-CONF_THRESH = 0.4
-IOU_THRESHOLD = 0.1
+CONF_THRESH = 0.75
+IOU_THRESHOLD = 0.4
 np.set_printoptions(threshold=np.inf)
 
 def plot_one_box(x, landmark,img, color=None, label=None, line_thickness=None):
@@ -180,9 +180,8 @@ def destroy(self):
 
     def preprocess_image(self, input_image_path):
         """
-        description: Read an image from image path, convert it to RGB,
-                     resize and pad it to target size, normalize to [0,1],
-                     transform to NCHW format.
+        description: Read an image from image path, resize and pad it to target size,
+                     normalize to [0,1],transform to NCHW format.
         param:
             input_image_path: str, image path
         return:
@@ -193,8 +192,7 @@ def preprocess_image(self, input_image_path):
         """
         image_raw = cv2.imread(input_image_path)
         h, w, c = image_raw.shape
-        image = cv2.cvtColor(image_raw, cv2.COLOR_BGR2RGB)
-        image = cv2.resize(image, (INPUT_W, INPUT_H))
+        image = cv2.resize(image_raw, (INPUT_W, INPUT_H))
 
         image = image.astype(np.float32)