Improve body tracking performance of VNDetectHumanBodyPoseRequest - ios

Question

I am trying to improve the performance of drawing the skeleton with body tracking for VNDetectHumanBodyPoseRequest even when further than 5 metres away, and with a stable iPhone XS camera.
The tracking has low confidence for the lower right limbs of my body, noticeable lag and there is jitter. I am unable to replicate the showcased performance in this year's WWDC demo video.
Here is the relevant code, adapted from Apple's sample code:
class Predictor {
  func extractPoses(_ sampleBuffer: CMSampleBuffer) throws -> [VNRecognizedPointsObservation] {
    let requestHandler = VNImageRequestHandler(cmSampleBuffer: sampleBuffer, orientation: .down)
    
    let request = VNDetectHumanBodyPoseRequest()
    
    do {
      // Perform the body pose-detection request.
      try requestHandler.perform([request])
    } catch {
      print("Unable to perform the request: \(error).\n")
    }
    
    return (request.results as? [VNRecognizedPointsObservation]) ?? [VNRecognizedPointsObservation]()
  }
}
I've captured the video data and am handling the sample buffers here:
class CameraViewController: AVCaptureVideoDataOutputSampleBufferDelegate {
func captureOutput(_ output: AVCaptureOutput, didOutput sampleBuffer: CMSampleBuffer, from connection: AVCaptureConnection) {
    let observations = try? predictor.extractPoses(sampleBuffer)
    observations?.forEach { processObservation($0) }
}
func processObservation(_ observation: VNRecognizedPointsObservation) {
    
    // Retrieve all torso points.
    guard let recognizedPoints =
            try? observation.recognizedPoints(forGroupKey: .all) else {
      return
    }
    
    let storedPoints = Dictionary(uniqueKeysWithValues: recognizedPoints.compactMap { (key, point) -> (String, CGPoint)? in
      return (key.rawValue, point.location)
    })
    
    DispatchQueue.main.sync {
      let mappedPoints = Dictionary(uniqueKeysWithValues: recognizedPoints.compactMap { (key, point) -> (String, CGPoint)? in
        guard point.confidence > 0.1 else { return nil }
        let norm = VNImagePointForNormalizedPoint(point.location,
                                                  Int(drawingView.bounds.width),
                                                  Int(drawingView.bounds.height))
        return (key.rawValue, norm)
      })
      
      let time = 1000 * observation.timeRange.start.seconds
      
      
      // Draw the points onscreen.
      DispatchQueue.main.async {
        self.drawingView.draw(points: mappedPoints)
      }
    }
  }
}
The drawingView.draw function is for a custom UIView on top of the camera view, and draws the points using CALayer sublayers. The AVCaptureSession code is exactly the same as the sample code here.
I tried using the VNDetectHumanBodyPoseRequest(completionHandler:) variant but this made no difference to the performance for me. I could try smoothing with a moving average filter though.. but there is still a problem with outlier predictions which are very inaccurate.
What am I missing?

Answer 1

This was a bug on iOS 14 beta v1-v3 I think. After upgrading to v4 and later beta releases, tracking is much better. The API also became a bit clearer with fine-grained type names with the latest beta updates.
Note I didn’t get an official answer from Apple regarding this bug, but this problem will probably completely disappear in the official iOS 14 release.