Tfjs working on older android 11 but not on 14

General Discussion
,
#1

We are using tfjs and pose-detection libs on Expo application and everything is working ok with older Android 11 (and ios) but Android 14 is failing, meaning pose detection return empty array without any error codes.

Are there any known issues using newer Android versions?

Some code for reference:

import { decodeJpeg } from '@tensorflow/tfjs-react-native'
import * as posedetection from '@tensorflow-models/pose-detection'
:
      await tf.ready()
      const detector = await posedetection.createDetector(
        posedetection.SupportedModels.BlazePose, { runtime: 'tfjs', }
      )
      const imageTensor = decodeJpeg(imageData)
      const poses = await detector.estimatePoses(imageTensor)
      console.log({ poses })
#3

An update on the imageTensor used in the example. The tensors are identical with android 11 & 14:

{“imageTensor”: {“dataId”: {“id”: 390}, “dtype”: “int32”, “id”: 390, “isDisposedInternal”: false, “kept”: false, “rankType”: “3”, “shape”: [720, 1280, 3], “size”: 2764800, “strides”: [3840, 3]}}

{“imageTensor”: {“dataId”: {“id”: 390}, “dtype”: “int32”, “id”: 390, “isDisposedInternal”: false, “kept”: false, “rankType”: “3”, “shape”: [720, 1280, 3], “size”: 2764800, “strides”: [3840, 3]}}

Seems that for some unknown reason pose estimation is not working on Android 14.

#4

Changing the detector model to posenet or movenet helps somewhat as poses are detected on Android 14 too. What is surprising is that the score is significantly better with older android using same setup:

Android 11:

LOG {“imageTensor”: {“dataId”: {“id”: 62}, “dtype”: “int32”, “id”: 62, “isDisposedInternal”: false, “kept”: false, “rankType”: “3”, “shape”: [720, 1280, 3], “size”: 2764800, “strides”: [3840, 3]}}
LOG {“poses”: [{“keypoints”: [Array], “score”: 0.8781241792089799}]}

Android 14:

LOG {“imageTensor”: {“dataId”: {“id”: 62}, “dtype”: “int32”, “id”: 62, “isDisposedInternal”: false, “kept”: false, “rankType”: “3”, “shape”: [720, 1280, 3], “size”: 2764800, “strides”: [3840, 3]}}
LOG {“poses”: [{“keypoints”: [Array], “score”: 0.02013262562618098}]}

#5

Created repro project, which can be found from: GitHub - expolli/pose-test: Barebones tfjs image pose detection.

Would appreciate it if someone could comment why app is working only on older devices.

Also created a bug for tfjs in github: Tfjs pose detection not working with newer Android devices (12, 13, 14) · Issue #8278 · tensorflow/tfjs · GitHub.