Merge pull request #2337 from HydrogenSulfate/add_optional_det

Add flexible configuration of PP-ShiTu detection model

deploy/python/predict_system.py

@@ -31,7 +31,14 @@ class SystemPredictor(object):
 
         self.config = config
         self.rec_predictor = RecPredictor(config)
-        self.det_predictor = DetPredictor(config)
+
+        if not config["Global"]["det_inference_model_dir"]:
+            logger.info(
+                f"Found 'Global.det_inference_model_dir' empty({config['Global']['det_inference_model_dir']}), so det_predictor is disabled"
+            )
+            self.det_predictor = None
+        else:
+            self.det_predictor = DetPredictor(config)
 
         assert 'IndexProcess' in config.keys(), "Index config not found ... "
         self.return_k = self.config['IndexProcess']['return_k']
@@ -92,7 +99,10 @@ class SystemPredictor(object):
     def predict(self, img):
         output = []
         # st1: get all detection results
-        results = self.det_predictor.predict(img)
+        if self.det_predictor:
+            results = self.det_predictor.predict(img)
+        else:
+            results = []
 
         # st2: add the whole image for recognition to improve recall
         results = self.append_self(results, img.shape)

docs/en/PPShiTu/PPShiTuV2_introduction.md

@@ -198,6 +198,21 @@ The final output is as follows.
 [{'bbox': [437, 71, 660, 728], 'rec_docs': '元气森林', 'rec_scores': 0.7740249}, {'bbox': [221, 72, 449, 701], 'rec_docs': '元气森林', 'rec_scores': 0.6950992}, {'bbox': [794, 104, 979, 652], 'rec_docs': '元气森林', 'rec_scores': 0.6305153}]
 ```
 
+The recognition process supports flexible configuration. Users can choose not to use the object detection model, but directly input a single whole image into the feature extraction model, and calculate the feature vector for subsequent retrieval, thereby reducing the time-consuming of the overall recognition process. It can be achieved by the script below
+```shell
+# Use the following command to use the GPU for whole-image prediction
+python3.7 python/predict_system.py -c configs/inference_general.yaml -o Global.det_inference_model_dir=None
+
+# Use the following command to use the CPU for whole-image prediction
+python3.7 python/predict_system.py -c configs/inference_general.yaml -o Global.use_gpu=False -o Global.det_inference_model_dir=None
+```
+
+The final output is as follows
+```log
+INFO: find 'Global.det_inference_model_dir' empty(), so det_predictor is disabled
+[{'bbox': [0, 0, 1200, 802], 'rec_docs': '元气森林', 'rec_scores': 0.5696486}]
+```
+
 #### 4.3.2 multi images prediction
 
 If you want to predict the images in the folder, you can directly modify the `Global.infer_imgs` field in the configuration file, or you can modify the corresponding configuration through the following -o parameter.

docs/zh_CN/deployment/PP-ShiTu/cpp.md

@@ -348,11 +348,25 @@ cd ..
   ./build/pp_shitu -c inference_drink.yaml
   ```
 
-  以 `drink_dataset_v2.0/test_images/nongfu_spring.jpeg` 作为输入图像，则执行上述推理命令可以得到如下结果
+  默认以 `../drink_dataset_v2.0/test_images/100.jpeg` 作为输入图像，则执行上述推理命令可以得到如下结果
 
   ```log
-  ../../deploy/drink_dataset_v2.0/test_images/nongfu_spring.jpeg:
-        result0: bbox[0, 0, 729, 1094], score: 0.688691, label: 农夫山泉-饮用天然水
+  ../drink_dataset_v2.0/test_images/100.jpeg:
+        result0: bbox[437, 72, 660, 723], score: 0.769916, label: 元气森林
+        result1: bbox[220, 71, 449, 685], score: 0.695485, label: 元气森林
+        result2: bbox[795, 104, 979, 653], score: 0.626963, label: 元气森林
+  ```
+
+  识别流程支持灵活配置，用户可以选择不使用主体检测模型，而直接将单幅整图输入到特征提取模型，计算特征向量供后续检索使用，从而减少整体识别流程的耗时。只需将`Global.det_inference_model_dir`后的字段改为`null`或者`""`，再运行以下推理命令即可
+  ```shell
+  ./build/pp_shitu -c inference_drink.yaml
+  ```
+
+  最终输出结果如下
+  ```log
+  Found 'Global.det_inference_model_dir' empty, so det_predictor is disabled
+  ../drink_dataset_v2.0/test_images/100.jpeg:
+          result0: bbox[0, 0, 1199, 801], score: 0.568903, label: 元气森林
   ```
 
   由于python和C++的opencv实现存在部分不同，可能导致python推理和C++推理结果有微小差异。但基本不影响最终的检索结果。

docs/zh_CN/quick_start/quick_start_recognition.md

@@ -223,6 +223,20 @@ python3.7 python/predict_system.py -c configs/inference_general.yaml -o Global.u
 
 ![](../../images/recognition/drink_data_demo/output/100.jpeg)
 
+识别流程支持灵活配置，用户可以选择不使用主体检测模型，而直接将单幅整图输入到特征提取模型，计算特征向量供后续检索使用，从而减少整体识别流程的耗时。可以按照以下命令直接进行整图识别
+```shell
+# 使用下面的命令使用 GPU 进行整图预测
+python3.7 python/predict_system.py -c configs/inference_general.yaml -o Global.det_inference_model_dir=None
+
+# 使用下面的命令使用 CPU 进行整图预测
+python3.7 python/predict_system.py -c configs/inference_general.yaml -o Global.use_gpu=False -o Global.det_inference_model_dir=None
+```
+
+最终输出结果如下
+```log
+INFO: Found 'Global.det_inference_model_dir' empty(), so det_predictor is disabled
+[{'bbox': [0, 0, 1200, 802], 'rec_docs': '元气森林', 'rec_scores': 0.5696486}]
+```
 
 <a name="基于文件夹的批量识别"></a>