是否有人在移动设备上成功运行posenet tflite?(Android或iOS)

时间:2019-05-31 06:02:55

标签: android ios tensorflow tensorflow-lite pose-estimation

Tensorflow提供了他们的姿势估计模型:multi_person_mobilenet_v1_075_float.tflitehosted_models

我发现此模型的输出已连接。生成的热图中的数字似乎是随机的。有人成功运行了吗?

我关注了博客:browser-with-tensorflow

我在下面附加了我的代码:

for(int i = 0; i < keyPointsNum; i ++) {
  for(int x = 0; x < outputW; x++){
    for(int y = 0; y < outputH; y++){
      heatMapArray[0][x][y][i] =  (float)(1.0 / ( Math.exp(heatMapArray[0][x][y][i] * -1) + 1));
    }
  }
}

// 2. argmax2d
int[][] keyPointsRaw = new int[keyPointsNum][2];
float[] scores = new float[keyPointsNum];
for(int i=0; i<keyPointsNum; i++){
  float maxScore = -10000f;
  for(int y = 0; y < outputW; y++){
    for(int x = 0; x < outputH; x++){
      if(maxScore < heatMapArray[0][y][x][i]){
        maxScore = heatMapArray[0][y][x][i];
        keyPointsRaw[i][X] = x;
        keyPointsRaw[i][Y] = y;
        scores[i] = maxScore;
      }
    }
  }
}

// 3. keypoint
float[][] keyPoints = new float[keyPointsNum][2];
for(int k=0; k<keyPointsNum; k++){
  float[] offsetVector = {offsets[0][keyPointsRaw[k][1]][keyPointsRaw[k][0]][k], offsets[0][keyPointsRaw[k][1]][keyPointsRaw[k][0]][k+17]};
  keyPoints[k][X] = keyPointsRaw[k][X] * 16 + offsetVector[X];
  keyPoints[k][Y] = keyPointsRaw[k][Y] * 16 + offsetVector[Y];
}

0 个答案:

没有答案