Merge branch 'dev'

Author: banderlog

README.md

@@ -75,7 +75,7 @@ For additional info read `cv2.getBuildInformation()` output.
 
 You will need ~7GB RAM and ~10GB disk space
 
-I am using Ubuntu 18.04 [multipass](https://multipass.run/) instance: `multipass launch -c 6 -d 10G -m 7G 18.04`.
+I am using Ubuntu 18.04 (python 3.6) [multipass](https://multipass.run/) instance: `multipass launch -c 6 -d 10G -m 7G 18.04`.
 
 ### Requirements
 
@@ -104,6 +104,7 @@ sudo ln -s  /usr/bin/python3 /usr/bin/python
 ```bash
 git clone https://github.com/banderlog/opencv-python-inference-engine
 cd opencv-python-inference-engine
+# git checkout dev
 ./download_all_stuff.sh
 ```
 

build/opencv/opencv_setup.sh

@@ -44,7 +44,7 @@ cmake -D CMAKE_BUILD_TYPE=RELEASE \
       -D FFMPEG_INCLUDE_DIRS=$FFMPEG_PATH/include \
       -D INF_ENGINE_INCLUDE_DIRS=$ABS_PORTION/dldt/inference-engine/include \
       -D INF_ENGINE_LIB_DIRS=$ABS_PORTION/dldt/bin/intel64/Release/lib \
-      -D INF_ENGINE_RELEASE=2021030000 \
+      -D INF_ENGINE_RELEASE=2021040000 \
       -D INSTALL_CREATE_DISTRIB=ON \
       -D INSTALL_C_EXAMPLES=OFF \
       -D INSTALL_PYTHON_EXAMPLES=OFF \

create_wheel/setup.py

@@ -15,7 +15,7 @@ def __len__(self):
 
 setuptools.setup(
     name='opencv-python-inference-engine',
-    version='2021.04.13',
+    version='2021.07.10',
     url="https://github.com/banderlog/opencv-python-inference-engine",
     maintainer="Kabakov Borys",
     license='MIT, Apache 2.0',

dldt

@@ -23,7 +23,8 @@ if test $(lsb_release -rs) != 18.04; then
 fi
 
 green "RESET GIT SUBMODULES"
-# use `git fetch --unshallow && git checkout tags/<tag>` for update
+# git checkout dev
+# use `git fetch --tags && git checkout tags/<tag>` for update
 git submodule update --init --recursive --depth=1 --jobs=4
 # restore changes command will differ between GIT versions (e.g., `restore`)
 git submodule foreach --recursive git checkout .
@@ -34,7 +35,6 @@ green "CLEAN BUILD DIRS"
 find build/dldt/ -mindepth 1 -not -name 'dldt_setup.sh' -not -name '*.patch' -delete
 find build/opencv/ -mindepth 1 -not -name 'opencv_setup.sh' -delete
 find build/ffmpeg/ -mindepth 1 -not -name 'ffmpeg_*.sh' -delete
-find build/openblas/ -mindepth 1 -not -name 'openblas_setup.sh' -delete
 
 green "CLEAN WHEEL DIR"
 find create_wheel/cv2/ -type f -not -name '__init__.py' -delete

download_all_stuff.sh

@@ -19,7 +19,6 @@ cd tests
 
 Something like below. The general idea is to test only inference speed, without preprocessing and decoding.
 Also, 1st inference must not count, because it will load all stuff into memory.
-I prefer to do such things in `ipython` or `jupyter` with `%timeit`.
 
 **NB:** be strict about Backend and Target
 

ffmpeg

@@ -68,30 +68,10 @@ for i in "${models[@]}"; do
         wget "${url_start}/${i%.*}/FP32/${i}"
     else
 	# checksum
-        sha256sum -c "${i}.sha256sum"
-    fi
-done
-
-
-# for speed test
-# {filename: file_google_drive_id}
-declare -A se_net=(["se_net.bin"]="1vbonFjVyleGRSd_wR-Khc1htsZybiHCG"
-		   ["se_net.xml"]="1Bz3EQwnes_iZ14iKAV6H__JZ2lynLmQz")
-
-# for each key
-for i in "${!se_net[@]}"; do
-    # if file exist
-    if [ -f $i ]; then
-        # checksum
-        sha256sum -c "${i}.sha256sum"
-    else
-        # get fileid from associative array and download file
-        wget --no-check-certificate "https://docs.google.com/uc?export=download&id=${se_net[$i]}" -O $i
+        sha256sum -c "${i}.sha256sum" || red "PROBLEMS ^^^"
     fi
 done
 
 green "For \"$WHEEL\""
 green "RUN TESTS with ./venv_t/bin/python ./tests.py"
 ./venv_t/bin/python ./tests.py
-green "RUN TESTS with ./venv_t/bin/python ./speed_test.py"
-./venv_t/bin/python ./speed_test.py

opencv

@@ -37,7 +37,7 @@ def _get_confidences(self, img: np.ndarray, box: tuple) -> np.ndarray:
         return outs
 
     def do_ocr(self, img: np.ndarray, bboxes: List[tuple]) -> List[str]:
-        """ Run OCR pipeline for a single words
+        """ Run OCR pipeline with greedy decoder for each single word (bbox)
 
             :param img: BGR image
             :param bboxes: list of sepaate word bboxes (ymin ,xmin ,ymax, xmax)
@@ -60,7 +60,7 @@ def do_ocr(self, img: np.ndarray, bboxes: List[tuple]) -> List[str]:
         for box in bboxes:
             # confidence distribution across symbols
             confs = self._get_confidences(img, box)
-            # get maximal confidence for the whole beam width
+            # get maximal confidence for the whole beam width aka greedy decoder
             idxs = confs[:, 0, :].argmax(axis=1)
             # drop blank characters '#' with id == 36 in charvec
             # isupposedly we taking only separate words as input