From cc2c61c7e4aa6d48d18b6e09da2566f6e9fdff98 Mon Sep 17 00:00:00 2001
From: Dmitry Kurtaev <dmitry.kurtaev+github@gmail.com>
Date: Mon, 29 Jan 2018 18:35:20 +0300
Subject: [PATCH] Updated TensorFlow Object Detection API (markdown)

---
 TensorFlow-Object-Detection-API.md | 56 +++++++++++++++++++++++++++++-
 1 file changed, 55 insertions(+), 1 deletion(-)

diff --git a/TensorFlow-Object-Detection-API.md b/TensorFlow-Object-Detection-API.md
index 3cb3080..dc6dc7b 100644
--- a/TensorFlow-Object-Detection-API.md
+++ b/TensorFlow-Object-Detection-API.md
@@ -57,6 +57,60 @@ cv.waitKey()
 ```
 ![](https://user-images.githubusercontent.com/25801568/35504975-a5db962c-04f5-11e8-9a9f-1d803f86af7f.png)
 
+## Run network in OpenCV
+OpenCV needs an extra configuration file to import object detection models from TensorFlow. It's based on a text version of the same serialized graph in protocol buffers format (protobuf).
+
+### Use existing config file for your model
+You can use one of the configs that has been tested in OpenCV. Choose it depends on your model and TensorFlow version:
+
+| Model | Version | ||
+|-------|-------------|----|----|
+| MobileNet-SSD | TensorFlow >= 1.4 | [weights](http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz) | |
+| Inception v2 SSD | TensorFlow >= 1.4 | [weights](http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz) | [config](https://github.com/opencv/opencv_extra/tree/master/testdata/dnn/ssd_inception_v2_coco_2017_11_17.pbtxt) |
+| MobileNet-SSD | TensorFlow < 1.4 | [weights](http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_11_06_2017.tar.gz) | [config](https://github.com/opencv/opencv_extra/blob/master/testdata/dnn/ssd_mobilenet_v1_coco.pbtxt) |
+
+
+
+### Get a text graph representation
+You can use the following script to make a text graph representation. It removes weights nodes and some unused fields.
+
+```python
+import tensorflow as tf
+
+# Read the graph.
+with tf.gfile.FastGFile('graph.pb') as f:
+    graph_def = tf.GraphDef()
+    graph_def.ParseFromString(f.read())
+
+# Remove Const nodes.
+for i in reversed(range(len(graph_def.node))):
+    if graph_def.node[i].op == 'Const':
+        del graph_def.node[i]
+    for attr in ['T', 'data_format', 'Tshape', 'N', 'Tidx', 'Tdim',
+                 'use_cudnn_on_gpu', 'Index', 'Tperm', 'is_training',
+                 'Tpaddings']:
+        if attr in graph_def.node[i].attr:
+            del graph_def.node[i].attr[attr]
+
+# Save as text.
+tf.train.write_graph(graph_def, "", "graph.pbtxt", as_text=True)
+```
+### Generate a config file
+Run `optimize_for_inference.py` tool to make your model simpler:
+```bash
+python ~/tensorflow/tensorflow/python/tools/optimize_for_inference.py \
+  --input frozen_inference_graph.pb \
+  --output opt_graph.pb \
+  --input_names image_tensor \
+  --output_names "num_detections,detection_scores,detection_boxes,detection_classes" \
+  --frozen_graph
+```
+
+Run a [graph transformation tool](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/graph_transforms/README.md#using-the-graph-transform-tool) to fuse constant nodes.
+
+Run [tf_text_graph.py]() script. If your model has different values of `num_classes`, `min_scale`, `max_scale`, `num_layers` or `aspect_ratios` comparing to [origin configuration files](https://github.com/tensorflow/models/tree/master/research/object_detection/samples/configs), specify it in the script arguments.
+
 ## References
 * [TensorFlow library](https://www.tensorflow.org/)
-* [COCO dataset](http://cocodataset.org/#home)
\ No newline at end of file
+* [COCO dataset](http://cocodataset.org/#home)
+* [Google Protobuf](https://developers.google.com/protocol-buffers/)
\ No newline at end of file