Add new TF Lite models to doc

PiperOrigin-RevId: 271248192

tensorflow/lite/g3doc/_book.yaml

@@ -110,25 +110,24 @@ upper_tabs:
       contents:
       - title: "Overview"
         path: /lite/models/
-      - heading: "Image classification"
-      - title: "Overview"
+      - heading: "Vision"
+      - title: "Image classification"
         path: /lite/models/image_classification/overview
-      - title: "Android"
-        path: /lite/models/image_classification/android
-      - title: "iOS"
-        path: /lite/models/image_classification/ios
-      - heading: "Other techniques"
+      - title: "Image segmentation"
+        path: /lite/models/segmentation/overview
       - title: "Object detection"
         path: /lite/models/object_detection/overview
       - title: "Pose estimation"
         path: /lite/models/pose_estimation/overview
-      - title: "Segmentation"
-        path: /lite/models/segmentation/overview
       - title: "Style transfer"
         path: /lite/models/style_transfer/overview
+      - title: "Face detection"
+        path: /lite/models/face_detection/overview
+      - title: "Hair segmentation"
+        path: /lite/models/hair_segmentation/overview
+      - heading: "Language"
       - title: "Smart reply"
         path: /lite/models/smart_reply/overview
-
     - name: "Python API"
       skip_translation: true
       contents:

tensorflow/lite/g3doc/models/face_detection/overview.md

@@ -0,0 +1,66 @@
+# Face detection
+
+## Get started
+
+BlazeFace is a lightweight face detection model, designed specifically for
+selfie use-case for mobile devices. It works for faces up to 2 meters from
+camera and provides 6 additional facial keypoints, which allows to estimate face
+rotation angles and do basic AR-effects on top of it.
+
+For a working demo of an ultrafast realtime face detection Android app using the
+model, check out this example by
+[MediaPipe](https://mediapipe.readthedocs.io/en/latest/):
+
+<a class="button button-primary" href="https://github.com/google/mediapipe/blob/master/mediapipe/docs/face_detection_mobile_gpu.md">Android
+example</a>
+<a class="button button-primary" href="https://github.com/google/mediapipe/raw/master/mediapipe/models/face_detection_front.tflite">Download
+starter model</a>
+
+### How it works
+
+BlazeFace is a lightweight and well-performing face detector tailored for mobile
+GPU inference. It runs at a speed of 200–1000+ FPS on flagship devices. This
+super-realtime performance enables it to be applied to any augmented reality
+pipeline that requires an accurate facial region of interest as an input for
+task-specific models, such as 2D/3D facial keypoint or geometry estimation,
+facial features or expression classification, and face region segmentation.
+
+The new techniques implemented in the model are:
+
+*   lightweight feature extraction network
+*   GPU-friendly anchor scheme
+*   improved tie resolution strategy alternative to non-maximum suppression
+
+Per each prediction, BlazeFace predicts face bounding box and 2D coordinates for
+6 facial keypoints:
+
+Id  | Part
+--- | -----------------
+0   | left_eye
+1   | right_eye
+2   | nose_tip
+3   | mouth_center
+4   | left_ear_tragion
+5   | right_ear_tragion
+
+### Examples of face detection
+
+![Demo](https://storage.googleapis.com/download.tensorflow.org/models/tflite/face_detection/demo.gif)
+
+### How it performs
+
+Model works in several predefined recommended resolutions, depending on input
+screen aspect ratio: **128x96**, **128x128**, **96x128**. For resolution
+**128x96** inference times shown below:
+
+Device                         | Inference time (ms)
+------------------------------ | -------------------
+Apple iPhone 7                 | 1.8
+Apple iPhone XS                | 0.6
+Google Pixel 3                 | 3.4
+Huawei P20                     | 5.8
+Samsung Galaxy S9+ (SM-G965U1) | 3.7
+
+### Read more about BlazeFace
+
+*   [Paper: BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs](https://sites.google.com/corp/view/perception-cv4arvr/blazeface)

tensorflow/lite/g3doc/models/hair_segmentation/overview.md

@@ -0,0 +1,42 @@
+# Hair segmentation
+
+## Get started
+
+Hair segmentation model produces a high-quality hair segmentation mask that is
+well suited for AR effects, e.g. virtual hair recoloring.
+
+For a working demo of a live hair recoloring Android app using the model, check
+out this example by [MediaPipe](https://mediapipe.readthedocs.io/en/latest/):
+
+<a class="button button-primary" href="https://github.com/google/mediapipe/blob/master/mediapipe/docs/hair_segmentation_mobile_gpu.md">Android
+example</a>
+<a class="button button-primary" href="https://github.com/google/mediapipe/raw/master/mediapipe/models/hair_segmentation.tflite">Download
+starter model</a>
+
+### How it works
+
+Hair segmentation refers to computer vision techniques that detect human hair in
+images and videos. To be clear, this technology is not recognizing who is in an
+image. The algorithm only estimates where is hair on an image and where is
+everything else.
+
+The model takes a video frame as input and returns a mask that tests if a pixel
+is a hair. For better results, this resulting mask is used as an additional
+input to the next frame.
+
+### Model architecture
+
+Standard hourglass segmentation network architecture with skip connections used
+for this model. The input is a 512x512x4 matrix. Channels are red, green, blue,
+previous mask or zeros for the first frame. The output is a 512x512x2 matrix
+with a background in the first channel and a hair mask in the second.
+
+![Model architecture](https://storage.googleapis.com/download.tensorflow.org/models/tflite/hair_segmentation/model_architecture.png)
+
+### Examples of hair recoloring
+
+![Sample](https://storage.googleapis.com/download.tensorflow.org/models/tflite/hair_segmentation/demo.gif)
+
+### Read more about hair segmentation
+
+*   [Paper: Real-time Hair segmentation and recoloring on Mobile GPUs](https://sites.google.com/corp/view/perception-cv4arvr/hair-segmentation)