Upload 4 files
Browse files- README.md +5 -5
- RunBlazeFace.cs +70 -57
- blazeface.sentis +0 -0
- info.json +2 -2
README.md
CHANGED
|
@@ -4,18 +4,18 @@ library_name: unity-sentis
|
|
| 4 |
pipeline_tag: object-detection
|
| 5 |
---
|
| 6 |
|
| 7 |
-
# Blaze Face face detector in Unity Sentis (Version 1.
|
| 8 |
-
*Version 1.3.0
|
| 9 |
|
| 10 |
This is the [Blaze Face model](https://developers.google.com/mediapipe/solutions/vision/face_detector) formatted to work in Unity Sentis 2023
|
| 11 |
|
| 12 |
## How to Use
|
| 13 |
* Create a new scene in Unity 2023
|
| 14 |
-
* Install `com.unity.sentis` version `1.
|
| 15 |
-
*
|
| 16 |
* Put a video in the Assets/StreamingAssets folder and set _videoName variable to the video name
|
| 17 |
* Create a RawImage and place it in your scene. Link to this image in the _previewUI field.
|
| 18 |
-
* Attach a sprite for the bounding box image to the
|
| 19 |
|
| 20 |
## Preview
|
| 21 |
When you get it working you should see something like this:
|
|
|
|
| 4 |
pipeline_tag: object-detection
|
| 5 |
---
|
| 6 |
|
| 7 |
+
# Blaze Face face detector in Unity Sentis (Version 1.4.0-pre.3*)
|
| 8 |
+
*Version 1.3.0 sentis files are not compatible with 1.4.0 and will need to be recreated/downloaded
|
| 9 |
|
| 10 |
This is the [Blaze Face model](https://developers.google.com/mediapipe/solutions/vision/face_detector) formatted to work in Unity Sentis 2023
|
| 11 |
|
| 12 |
## How to Use
|
| 13 |
* Create a new scene in Unity 2023
|
| 14 |
+
* Install `com.unity.sentis` version `1.4.0-pre.3` from the package manager
|
| 15 |
+
* Drag the blazeface.sentis file into the model asset field
|
| 16 |
* Put a video in the Assets/StreamingAssets folder and set _videoName variable to the video name
|
| 17 |
* Create a RawImage and place it in your scene. Link to this image in the _previewUI field.
|
| 18 |
+
* Attach a sprite or texture for the bounding box image to the BondingBoxSprite or BorderTexture field
|
| 19 |
|
| 20 |
## Preview
|
| 21 |
When you get it working you should see something like this:
|
RunBlazeFace.cs
CHANGED
|
@@ -3,6 +3,7 @@ using Unity.Sentis;
|
|
| 3 |
using UnityEngine.Video;
|
| 4 |
using UnityEngine.UI;
|
| 5 |
using Lays = Unity.Sentis.Layers;
|
|
|
|
| 6 |
|
| 7 |
/*
|
| 8 |
* Blaze Face Inference
|
|
@@ -11,7 +12,7 @@ using Lays = Unity.Sentis.Layers;
|
|
| 11 |
* Basic inference script for blaze face
|
| 12 |
*
|
| 13 |
* Put this script on the Main Camera
|
| 14 |
-
*
|
| 15 |
* Create a RawImage in the scene
|
| 16 |
* Put a link to that image in previewUI
|
| 17 |
* Put a video in Assets/StreamingAssets folder and put the name of it int videoName
|
|
@@ -22,11 +23,15 @@ using Lays = Unity.Sentis.Layers;
|
|
| 22 |
|
| 23 |
public class RunBlazeFace : MonoBehaviour
|
| 24 |
{
|
|
|
|
|
|
|
|
|
|
| 25 |
//Drag a link to a raw image here:
|
| 26 |
public RawImage previewUI = null;
|
| 27 |
|
| 28 |
// Put your bounding box sprite image here
|
| 29 |
-
public Sprite
|
|
|
|
| 30 |
|
| 31 |
// 6 optional sprite images (left eye, right eye, nose, mouth, left ear, right ear)
|
| 32 |
public Sprite[] markerTextures;
|
|
@@ -58,9 +63,6 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 58 |
//Holds input image size
|
| 59 |
int size = 128;
|
| 60 |
|
| 61 |
-
Ops ops;
|
| 62 |
-
ITensorAllocator allocator;
|
| 63 |
-
|
| 64 |
Model model;
|
| 65 |
|
| 66 |
//webcam device name:
|
|
@@ -68,6 +70,8 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 68 |
|
| 69 |
bool closing = false;
|
| 70 |
|
|
|
|
|
|
|
| 71 |
public struct BoundingBox
|
| 72 |
{
|
| 73 |
public float centerX;
|
|
@@ -78,7 +82,6 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 78 |
|
| 79 |
void Start()
|
| 80 |
{
|
| 81 |
-
allocator = new TensorCachingAllocator();
|
| 82 |
|
| 83 |
//(Note: if using a webcam on mobile get permissions here first)
|
| 84 |
|
|
@@ -88,7 +91,7 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 88 |
|
| 89 |
SetupModel();
|
| 90 |
|
| 91 |
-
|
| 92 |
}
|
| 93 |
|
| 94 |
void SetupInput()
|
|
@@ -120,6 +123,16 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 120 |
}
|
| 121 |
}
|
| 122 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
void Update()
|
| 124 |
{
|
| 125 |
if (inputType == InputType.Webcam)
|
|
@@ -198,36 +211,43 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 198 |
{
|
| 199 |
float[] offsets = GetGridBoxCoords();
|
| 200 |
|
| 201 |
-
model = ModelLoader.Load(
|
| 202 |
|
| 203 |
//We need to add extra layers to the model in order to aggregate the box predicions:
|
| 204 |
|
| 205 |
size = model.inputs[0].shape.ToTensorShape()[1]; // Input tensor width
|
| 206 |
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 231 |
}
|
| 232 |
|
| 233 |
void DrawFaces(TensorFloat index3, TensorFloat regressors, int NMAX, Vector2 scale)
|
|
@@ -237,12 +257,12 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 237 |
//Draw bounding box of face
|
| 238 |
var box = new BoundingBox
|
| 239 |
{
|
| 240 |
-
centerX = index3[
|
| 241 |
-
centerY = index3[
|
| 242 |
-
width = index3[
|
| 243 |
-
height = index3[
|
| 244 |
};
|
| 245 |
-
DrawBox(box,
|
| 246 |
if (regressors == null) continue;
|
| 247 |
|
| 248 |
//Draw markers for eyes, ears, nose, mouth:
|
|
@@ -250,12 +270,12 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 250 |
{
|
| 251 |
var marker = new BoundingBox
|
| 252 |
{
|
| 253 |
-
centerX = box.centerX + (regressors[
|
| 254 |
-
centerY = box.centerY + (regressors[
|
| 255 |
width = 8.0f * scale.x,
|
| 256 |
height = 8.0f * scale.y,
|
| 257 |
};
|
| 258 |
-
DrawBox(marker, j < markerTextures.Length ? markerTextures[j] :
|
| 259 |
}
|
| 260 |
}
|
| 261 |
}
|
|
@@ -265,31 +285,24 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 265 |
var transform = new TextureTransform();
|
| 266 |
transform.SetDimensions(size, size, 3);
|
| 267 |
transform.SetTensorLayout(0, 3, 1, 2);
|
| 268 |
-
using var
|
| 269 |
-
|
| 270 |
-
// Pre-process the image to make input in range (-1..1)
|
| 271 |
-
using var image = ops.Mad(image0, 2f, -1f);
|
| 272 |
|
| 273 |
worker.Execute(image);
|
| 274 |
|
| 275 |
-
using var
|
| 276 |
-
using var
|
| 277 |
-
|
| 278 |
-
var NM1 = worker.PeekOutput("NMS") as TensorInt;
|
| 279 |
-
|
| 280 |
-
using var boxCoords2 = ops.Reshape(boxCoords, boxCoords.shape.Unsqueeze(0));
|
| 281 |
-
using var output = ops.GatherND(boxCoords2, NM1, 0);
|
| 282 |
|
| 283 |
-
|
| 284 |
-
|
| 285 |
|
| 286 |
-
|
| 287 |
-
|
| 288 |
|
| 289 |
ClearAnnotations();
|
| 290 |
|
| 291 |
Vector2 markerScale = previewUI.rectTransform.rect.size / size;
|
| 292 |
|
|
|
|
| 293 |
DrawFaces(output, markersOutput, output.shape[0], markerScale);
|
| 294 |
|
| 295 |
}
|
|
@@ -329,8 +342,8 @@ public class RunBlazeFace : MonoBehaviour
|
|
| 329 |
void CleanUp()
|
| 330 |
{
|
| 331 |
closing = true;
|
| 332 |
-
|
| 333 |
-
|
| 334 |
if (webcam) Destroy(webcam);
|
| 335 |
if (video) Destroy(video);
|
| 336 |
RenderTexture.active = null;
|
|
|
|
| 3 |
using UnityEngine.Video;
|
| 4 |
using UnityEngine.UI;
|
| 5 |
using Lays = Unity.Sentis.Layers;
|
| 6 |
+
using FF = Unity.Sentis.Functional;
|
| 7 |
|
| 8 |
/*
|
| 9 |
* Blaze Face Inference
|
|
|
|
| 12 |
* Basic inference script for blaze face
|
| 13 |
*
|
| 14 |
* Put this script on the Main Camera
|
| 15 |
+
* Drawg the blazeface.sentis into the modelAsset field
|
| 16 |
* Create a RawImage in the scene
|
| 17 |
* Put a link to that image in previewUI
|
| 18 |
* Put a video in Assets/StreamingAssets folder and put the name of it int videoName
|
|
|
|
| 23 |
|
| 24 |
public class RunBlazeFace : MonoBehaviour
|
| 25 |
{
|
| 26 |
+
//Draw the sentis model asset here:
|
| 27 |
+
public ModelAsset modelAsset;
|
| 28 |
+
|
| 29 |
//Drag a link to a raw image here:
|
| 30 |
public RawImage previewUI = null;
|
| 31 |
|
| 32 |
// Put your bounding box sprite image here
|
| 33 |
+
public Sprite boundingboxSprite;
|
| 34 |
+
public Texture2D borderTexture;
|
| 35 |
|
| 36 |
// 6 optional sprite images (left eye, right eye, nose, mouth, left ear, right ear)
|
| 37 |
public Sprite[] markerTextures;
|
|
|
|
| 63 |
//Holds input image size
|
| 64 |
int size = 128;
|
| 65 |
|
|
|
|
|
|
|
|
|
|
| 66 |
Model model;
|
| 67 |
|
| 68 |
//webcam device name:
|
|
|
|
| 70 |
|
| 71 |
bool closing = false;
|
| 72 |
|
| 73 |
+
TensorFloat anchors, centersToCorners;
|
| 74 |
+
|
| 75 |
public struct BoundingBox
|
| 76 |
{
|
| 77 |
public float centerX;
|
|
|
|
| 82 |
|
| 83 |
void Start()
|
| 84 |
{
|
|
|
|
| 85 |
|
| 86 |
//(Note: if using a webcam on mobile get permissions here first)
|
| 87 |
|
|
|
|
| 91 |
|
| 92 |
SetupModel();
|
| 93 |
|
| 94 |
+
SetupBoundingBoxSprite();
|
| 95 |
}
|
| 96 |
|
| 97 |
void SetupInput()
|
|
|
|
| 123 |
}
|
| 124 |
}
|
| 125 |
|
| 126 |
+
void SetupBoundingBoxSprite()
|
| 127 |
+
{
|
| 128 |
+
if (boundingboxSprite == null)
|
| 129 |
+
{
|
| 130 |
+
boundingboxSprite = Sprite.Create(borderTexture,
|
| 131 |
+
new Rect(0, 0, borderTexture.width, borderTexture.height),
|
| 132 |
+
new Vector2(borderTexture.width / 2, borderTexture.height / 2));
|
| 133 |
+
}
|
| 134 |
+
}
|
| 135 |
+
|
| 136 |
void Update()
|
| 137 |
{
|
| 138 |
if (inputType == InputType.Webcam)
|
|
|
|
| 211 |
{
|
| 212 |
float[] offsets = GetGridBoxCoords();
|
| 213 |
|
| 214 |
+
var model = ModelLoader.Load(modelAsset);
|
| 215 |
|
| 216 |
//We need to add extra layers to the model in order to aggregate the box predicions:
|
| 217 |
|
| 218 |
size = model.inputs[0].shape.ToTensorShape()[1]; // Input tensor width
|
| 219 |
|
| 220 |
+
anchors = new TensorFloat(new TensorShape(offsets.Length / 4, 4), offsets);
|
| 221 |
+
|
| 222 |
+
centersToCorners = new TensorFloat(new TensorShape(4, 4),
|
| 223 |
+
new float[]
|
| 224 |
+
{
|
| 225 |
+
1, 0, 1, 0,
|
| 226 |
+
0, 1, 0, 1,
|
| 227 |
+
-0.5f, 0, 0.5f, 0,
|
| 228 |
+
0, -0.5f, 0, 0.5f
|
| 229 |
+
});
|
| 230 |
+
|
| 231 |
+
var model2 = Functional.Compile(
|
| 232 |
+
input =>
|
| 233 |
+
{
|
| 234 |
+
var outputs = model.Forward(input);
|
| 235 |
+
var regressors = outputs[0][0]; //shape=(896,16)
|
| 236 |
+
var scores = outputs[1][0].Transpose(0, 1) - scoreThreshold; //shape=(1,896)
|
| 237 |
+
var boxCoords = regressors[.., 0..4] + FunctionalTensor.FromTensor(anchors);//(896,4)
|
| 238 |
+
var boxCorners = FF.MatMul(boxCoords, FunctionalTensor.FromTensor(centersToCorners));
|
| 239 |
+
var indices = FF.NMS(boxCorners, scores, iouThreshold); //shape=(N)
|
| 240 |
+
var indices2 = indices.Unsqueeze(-1).BroadcastTo(new int[] { 4 }); //shape=(N,4)
|
| 241 |
+
var output = FF.Gather(boxCoords, 0, indices2); //shape=(N,4)
|
| 242 |
+
var indices3 = indices.Unsqueeze(-1).BroadcastTo(new int[] { 16 }); //shape=(N,16)
|
| 243 |
+
var markersOutput = FF.Gather(regressors, 0, indices3); //shape=(N,16)
|
| 244 |
+
return (output, markersOutput);
|
| 245 |
+
},
|
| 246 |
+
InputDef.FromModel(model)[0]
|
| 247 |
+
);
|
| 248 |
+
|
| 249 |
+
worker = WorkerFactory.CreateWorker(backend, model2);
|
| 250 |
+
|
| 251 |
}
|
| 252 |
|
| 253 |
void DrawFaces(TensorFloat index3, TensorFloat regressors, int NMAX, Vector2 scale)
|
|
|
|
| 257 |
//Draw bounding box of face
|
| 258 |
var box = new BoundingBox
|
| 259 |
{
|
| 260 |
+
centerX = index3[ n, 0] * scale.x,
|
| 261 |
+
centerY = index3[ n, 1] * scale.y,
|
| 262 |
+
width = index3[ n, 2] * scale.x,
|
| 263 |
+
height = index3[ n, 3] * scale.y
|
| 264 |
};
|
| 265 |
+
DrawBox(box, boundingboxSprite);
|
| 266 |
if (regressors == null) continue;
|
| 267 |
|
| 268 |
//Draw markers for eyes, ears, nose, mouth:
|
|
|
|
| 270 |
{
|
| 271 |
var marker = new BoundingBox
|
| 272 |
{
|
| 273 |
+
centerX = box.centerX + (regressors[ n, 4 + j * 2] - regressors[ n, 0]) * scale.x,
|
| 274 |
+
centerY = box.centerY + (regressors[ n, 4 + j * 2 + 1] - regressors[ n, 1]) * scale.y,
|
| 275 |
width = 8.0f * scale.x,
|
| 276 |
height = 8.0f * scale.y,
|
| 277 |
};
|
| 278 |
+
DrawBox(marker, j < markerTextures.Length ? markerTextures[j] : boundingboxSprite);
|
| 279 |
}
|
| 280 |
}
|
| 281 |
}
|
|
|
|
| 285 |
var transform = new TextureTransform();
|
| 286 |
transform.SetDimensions(size, size, 3);
|
| 287 |
transform.SetTensorLayout(0, 3, 1, 2);
|
| 288 |
+
using var image = TextureConverter.ToTensor(source, transform);
|
|
|
|
|
|
|
|
|
|
| 289 |
|
| 290 |
worker.Execute(image);
|
| 291 |
|
| 292 |
+
using var output = worker.PeekOutput("output_0") as TensorFloat; //face coords
|
| 293 |
+
using var markersOutput = worker.PeekOutput("output_1") as TensorFloat; //contains markers
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 294 |
|
| 295 |
+
output.CompleteOperationsAndDownload();
|
| 296 |
+
markersOutput.CompleteOperationsAndDownload();
|
| 297 |
|
| 298 |
+
// Debug.Log(output.shape + " " + markersOutput.shape);
|
| 299 |
+
//return;
|
| 300 |
|
| 301 |
ClearAnnotations();
|
| 302 |
|
| 303 |
Vector2 markerScale = previewUI.rectTransform.rect.size / size;
|
| 304 |
|
| 305 |
+
//Debug.Log(output.shape + " " + markersOutput.shape);
|
| 306 |
DrawFaces(output, markersOutput, output.shape[0], markerScale);
|
| 307 |
|
| 308 |
}
|
|
|
|
| 342 |
void CleanUp()
|
| 343 |
{
|
| 344 |
closing = true;
|
| 345 |
+
anchors?.Dispose();
|
| 346 |
+
centersToCorners?.Dispose();
|
| 347 |
if (webcam) Destroy(webcam);
|
| 348 |
if (video) Destroy(video);
|
| 349 |
RenderTexture.active = null;
|
blazeface.sentis
CHANGED
|
Binary files a/blazeface.sentis and b/blazeface.sentis differ
|
|
|
info.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"code": [
|
| 3 |
-
|
| 4 |
],
|
| 5 |
"models": [
|
| 6 |
"blazeface.sentis"
|
| 7 |
],
|
| 8 |
"version": [
|
| 9 |
-
|
| 10 |
]
|
| 11 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"code": [
|
| 3 |
+
"RunBlazeFace.cs"
|
| 4 |
],
|
| 5 |
"models": [
|
| 6 |
"blazeface.sentis"
|
| 7 |
],
|
| 8 |
"version": [
|
| 9 |
+
"1.4.0"
|
| 10 |
]
|
| 11 |
}
|