[pose-detection]Use f16 model. (#708)

FEATURE

Author: lina128

pose-detection/src/blazepose_tfjs/blazepose_test.ts

@@ -24,7 +24,7 @@ import {expectArraysClose} from '@tensorflow/tfjs-core/dist/test_util';
 import * as poseDetection from '../index';
 import {getXYPerFrame, KARMA_SERVER, loadImage, loadVideo} from '../test_util';
 
-const EPSILON_IMAGE = 18;
+const EPSILON_IMAGE = 19;
 const EPSILON_VIDEO = 15;
 
 // ref:

pose-detection/src/blazepose_tfjs/constants.ts

@@ -18,13 +18,13 @@
 import {BlazePoseTfjsModelConfig} from './types';
 
 export const DEFAULT_BLAZEPOSE_DETECTOR_MODEL_URL =
-    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/detector/heatmap/model.json';
+    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/detector/f16/model.json';
 export const DEFAULT_BLAZEPOSE_LANDMARK_MODEL_URL_FULL =
-    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/full/model.json';
+    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/full-f16/model.json';
 export const DEFAULT_BLAZEPOSE_LANDMARK_MODEL_URL_LITE =
-    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/lite/model.json';
+    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/lite-f16/model.json';
 export const DEFAULT_BLAZEPOSE_LANDMARK_MODEL_URL_HEAVY =
-    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/heavy/model.json';
+    'https://storage.googleapis.com/tfjs-models/savedmodel/blazepose/landmark/heavy-f16/model.json';
 export const BLAZEPOSE_DETECTOR_ANCHOR_CONFIGURATION = {
   reduceBoxesInLowestlayer: false,
   interpolatedScaleAspectRatio: 1.0,

pose-detection/src/blazepose_tfjs/detector.ts

@@ -324,18 +324,18 @@ export class BlazePoseTfjsDetector extends BasePoseDetector {
     // Output[3]: This tensor (shape: [1, 195]) represents 39 5-d keypoints.
     // The first 33 refer to the keypoints. The final 6 key points refer to
     // the alignment points from the detector model and the hands.)
-    // Output [0]: This tensor (shape: [1, 1]) represents the confidence
+    // Output [4]: This tensor (shape: [1, 1]) represents the confidence
     // score.
-    // Output [2]: This tensor (shape: [1, 64, 64, 39]) represents heatmap for
+    // Output [1]: This tensor (shape: [1, 64, 64, 39]) represents heatmap for
     // the 39 landmarks.
     // Lite model:
-    // Output[1]: This tensor (shape: [1, 195]) represents 39 5-d keypoints.
-    // Output[2]: This tensor (shape: [1, 1]) represents the confidence score.
-    // Output[4]: This tensor (shape: [1, 64, 64, 39]) represents heatmap for
+    // Output[4]: This tensor (shape: [1, 195]) represents 39 5-d keypoints.
+    // Output[3]: This tensor (shape: [1, 1]) represents the confidence score.
+    // Output[1]: This tensor (shape: [1, 64, 64, 39]) represents heatmap for
     // the 39 landmarks.
     // Heavy model:
     // Output[3]: This tensor (shape: [1, 195]) represents 39 5-d keypoints.
-    // Output[2]: This tensor (shape: [1, 1]) represents the confidence score.
+    // Output[1]: This tensor (shape: [1, 1]) represents the confidence score.
     // Output[4]: This tensor (shape: [1, 64, 64, 39]) represents heatmap for
     // the 39 landmarks.
     const landmarkResult =
@@ -345,18 +345,18 @@ export class BlazePoseTfjsDetector extends BasePoseDetector {
 
     switch (this.modelType) {
       case 'lite':
-        landmarkTensor = landmarkResult[1] as tf.Tensor2D;
-        poseFlagTensor = landmarkResult[2] as tf.Tensor2D;
-        heatmapTensor = landmarkResult[4] as tf.Tensor4D;
+        landmarkTensor = landmarkResult[3] as tf.Tensor2D;
+        poseFlagTensor = landmarkResult[4] as tf.Tensor2D;
+        heatmapTensor = landmarkResult[1] as tf.Tensor4D;
         break;
       case 'full':
-        landmarkTensor = landmarkResult[3] as tf.Tensor2D;
-        poseFlagTensor = landmarkResult[0] as tf.Tensor2D;
-        heatmapTensor = landmarkResult[2] as tf.Tensor4D;
+        landmarkTensor = landmarkResult[4] as tf.Tensor2D;
+        poseFlagTensor = landmarkResult[3] as tf.Tensor2D;
+        heatmapTensor = landmarkResult[1] as tf.Tensor4D;
         break;
       case 'heavy':
         landmarkTensor = landmarkResult[3] as tf.Tensor2D;
-        poseFlagTensor = landmarkResult[2] as tf.Tensor2D;
+        poseFlagTensor = landmarkResult[1] as tf.Tensor2D;
         heatmapTensor = landmarkResult[4] as tf.Tensor4D;
         break;
       default: