Forum Archive

Hi,

I have studied face_detector.py from here.

Here detector is applied on video camera frames in real time.

The issue comes after 80-90 frames - the script simply stops, but the application didn't crash.

And I can't figure out why it stops and how I can prevent this?

Any ideas?

# coding: utf-8

from objc_util import *
from ctypes import c_void_p
import ui
import time

# 全フレームを処理しようとすると動かなくなるのでこの程度で
FRAME_INTERVAL = 6  # 30fps / 6 = 5fps

frame_counter = 0
last_fps_time = time.time()
fps_counter = 0

AVCaptureSession = ObjCClass('AVCaptureSession')
AVCaptureDevice = ObjCClass('AVCaptureDevice')
AVCaptureDeviceInput = ObjCClass('AVCaptureDeviceInput')
AVCaptureVideoDataOutput = ObjCClass('AVCaptureVideoDataOutput')
AVCaptureVideoPreviewLayer = ObjCClass('AVCaptureVideoPreviewLayer')

CIImage    = ObjCClass('CIImage')
CIDetector = ObjCClass('CIDetector')

dispatch_get_current_queue = c.dispatch_get_current_queue
dispatch_get_current_queue.restype = c_void_p

CMSampleBufferGetImageBuffer = c.CMSampleBufferGetImageBuffer
CMSampleBufferGetImageBuffer.argtypes = [c_void_p]
CMSampleBufferGetImageBuffer.restype = c_void_p

CVPixelBufferLockBaseAddress = c.CVPixelBufferLockBaseAddress
CVPixelBufferLockBaseAddress.argtypes = [c_void_p, c_int]
CVPixelBufferLockBaseAddress.restype = None

CVPixelBufferGetWidth = c.CVPixelBufferGetWidth
CVPixelBufferGetWidth.argtypes = [c_void_p]
CVPixelBufferGetWidth.restype = c_int

CVPixelBufferGetHeight = c.CVPixelBufferGetHeight
CVPixelBufferGetHeight.argtypes = [c_void_p]
CVPixelBufferGetHeight.restype = c_int

CVPixelBufferUnlockBaseAddress = c.CVPixelBufferUnlockBaseAddress
CVPixelBufferUnlockBaseAddress.argtypes = [c_void_p, c_int]
CVPixelBufferUnlockBaseAddress.restype = None


def captureOutput_didOutputSampleBuffer_fromConnection_(_self, _cmd, _output, _sample_buffer, _conn):
    global frame_counter, fps_counter, last_fps_time
    global image_width, image_height, faces

    # 性能確認のためビデオデータの実 FPS 表示
    fps_counter += 1
    now = time.time()
    if int(now) > int(last_fps_time):
        label_fps.text = '{:5.2f} fps'.format((fps_counter) / (now - last_fps_time))
        last_fps_time = now
        fps_counter = 0

    # 画像処理は FRAME_INTERVAL 間隔で処理
    if frame_counter == 0:
        # ビデオ画像のフレームデータを取得
        imagebuffer =  CMSampleBufferGetImageBuffer(_sample_buffer)
        # バッファをロック
        CVPixelBufferLockBaseAddress(imagebuffer, 0)

        image_width  = CVPixelBufferGetWidth(imagebuffer)
        image_height = CVPixelBufferGetHeight(imagebuffer)
        ciimage = CIImage.imageWithCVPixelBuffer_(ObjCInstance(imagebuffer))

        # CIDetector により顔検出
        options = {'CIDetectorAccuracy': 'CIDetectorAccuracyHigh'}
        detector = CIDetector.detectorOfType_context_options_('CIDetectorTypeFace', None, options)
        faces = detector.featuresInImage_(ciimage)

        # バッファのロックを解放
        CVPixelBufferUnlockBaseAddress(imagebuffer, 0)

        # 検出した顔の情報を使って表示を更新
        path_view.set_needs_display()

    frame_counter = (frame_counter + 1) % FRAME_INTERVAL

class PathView(ui.View):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def draw(self):
        # 検出した顔の輪郭に合わせて、表示を加工
        if faces is not None and faces.count() != 0:
            # 顔の部分を白く覆う
            ui.set_color((1, 1, 1, 0.9))
            for face in faces:
                face_bounds = face.bounds()
                # カメラの画像は X軸=1920 Y軸=1080
                # View は X軸=375 Y軸=667
                # 画像のX軸Y軸をViewのY軸X軸に対応させ、サイズを調整
                x = face_bounds.origin.y    * self.height / image_width
                y = face_bounds.origin.x    * self.width  / image_height
                w = face_bounds.size.height * self.height / image_width
                h = face_bounds.size.width  * self.width  / image_height
                path = ui.Path.oval(x, y, w * 1.3, h)
                path.fill()

@on_main_thread
def main():
    global path_view, label_fps, faces

    # 画面の回転には対応しておらず
    # iPhoneの画面縦向きでロックした状態で、横長画面で使う想定
    # View のサイズは手持ちの iPhone6 に合わせたもの
    faces = None
    main_view = ui.View(frame=(0, 0, 375, 667))
    path_view = PathView(frame=main_view.frame)
    main_view.name = 'Face Detector'

    sampleBufferDelegate = create_objc_class(
                                'sampleBufferDelegate',
                                methods=[captureOutput_didOutputSampleBuffer_fromConnection_],
                                protocols=['AVCaptureVideoDataOutputSampleBufferDelegate'])
    delegate = sampleBufferDelegate.new()

    session = AVCaptureSession.alloc().init()
    device = AVCaptureDevice.defaultDeviceWithMediaType_('vide')
    _input = AVCaptureDeviceInput.deviceInputWithDevice_error_(device, None)
    if _input:
        session.addInput_(_input)
    else:
        print('Failed to create input')
        return

    output = AVCaptureVideoDataOutput.alloc().init()
    queue = ObjCInstance(dispatch_get_current_queue())
    output.setSampleBufferDelegate_queue_(delegate, queue)
    output.alwaysDiscardsLateVideoFrames = True

    session.addOutput_(output)
    session.sessionPreset = 'AVCaptureSessionPresetHigh' # 1920 x 1080

    prev_layer = AVCaptureVideoPreviewLayer.layerWithSession_(session)
    prev_layer.frame = ObjCInstance(main_view).bounds()
    prev_layer.setVideoGravity_('AVLayerVideoGravityResizeAspectFill')

    ObjCInstance(main_view).layer().addSublayer_(prev_layer)

    # 性能確認のためビデオデータの実 FPS 表示
    label_fps = ui.Label(frame=(0, 0, main_view.width, 30), flex='W', name='fps')
    label_fps.background_color = (0, 0, 0, 0.5)
    label_fps.text_color = 'white'
    label_fps.text = ''
    label_fps.alignment = ui.ALIGN_CENTER

    main_view.add_subview(label_fps)
    main_view.add_subview(path_view)

    session.startRunning()

    main_view.present('sheet')
    main_view.wait_modal()

    session.stopRunning()
    delegate.release()
    session.release()
    output.release()

if __name__ == '__main__':
    main()

You might consider adding a way for frame_counter and last frame time to be periodically reset to 0 if frame_xounter ever gets to be large.

Or, have a label that gets updated with frame time, frame_counter , and now time -- since these are used in the logic inside the callback.

Did you test? Same here?

@pavlinb, works as-is for me, iPhone 11 Pro.

So, the fps measurement at the top didn’t hang after 80-90 seconds?

Do you fix the problem ?
because this is my problem too

@pavlinb, earlier you said 80-90 frames, which would be about 3 seconds. I have not yet let it run for minutes.

@pavlinb, if this was originally run on iPhone 6, I bet we can do better than 5 fps on the face detection today. Have you studied enough to know whether we really need to initialize the face detector with options every time it is used?

@mikael My goal is not face detection. I need the method to obtain frames from video camera and use them continuously.

@mikael Correct duration should be 80-90 passes of face detector.

When phone wakes up after auto lock, script makes another 80-90 passes and hangs again.

I wonder if you need an autorelease pool. There was a thread on this recently, there is a context manager in objc_util.

Also, get rid of wait_modal, but instead stop everything in a custom view will_close method. Wait_modal keeps your background thread from doing anything else.

@JonB Tried, but same behaviour.

Other way of obtaining video camera frames continuously is welcome.

For data processing or just displaying as a preview?

@JonB For data processing.

ok, i do see some issues..

first, we need to implement the didDropSampleBuffer delegate method, and display the reason for a late frame.

second, we need to set the minFrameDuration, so that the delegate is never called more than you need it. on older versions, i believe this was in output.minFrameDuration. in newer iOS versions, i believe you set on the connection, output.connections[0].videoMinFrameDuration

third, there is an issue regarding what dispatch queue things are called on. or, the delegate needs to always return as fast as possible, calling heavy lifting on another thread, and dropping data otherwise. i think what is happening is that your delegate gets called on a background thread, and just keeps falling farther and farther behind. it might be better to ensure the delegate is called on the main thread.

ill post a gist later with some improvements.

@JonB I have tested another good example here

In this script is implemented

self.captureOutput.setMinFrameDuration_(CMTimeMake(1, 2), argtypes=[CMTime], restype=None)

I have plaeyd with CMTimeMake(.,.).

Unfotunately this script also hangs.

I have a version based of Cethric's. I'll clean it up and post later. I've implemented the drop frame callback so you can see what the issues are.

One problem I have found is that the various ways to set minFrameDuration don't work. That means the callback gets called at a high rate.

It seems that setMinFrameDuration is in camera.captureOutput object ('AVCaptureVideoDataOutput').

But there is also setActiveVideoMinFrameDuration in camera.captureDevice ('AVCaptureDevice').

I couldn't make them working too both.

try this.
https://gist.github.com/424d4fe1a3c0b1ae3fd705d72f665c1e

increase FRAME_PROC_INTERVAL until you stop seeing FrameLate constantly. or set to 1 to go as fast as possible. you will probably see FrameLate, but if you start seeing OutOfBuffers or something like that, it means you are doing some processing and forgetting to release the capture buffer.

there are many hoops to jump through to get the actual min frame interval set, i think i have all the steps correct, though am on an older device. see if it works for you. I have a DESIRED_FPS that you can change to see if you can get <30 fps..

does this one hang? if so, what message do you get on the first frame when it comes back?

This one works: link

from ctypes import *
from objc_util import *

class CMTime(Structure):
  _fields_ = [('value', c_int64), ('timescale', c_int32), ('flags', c_uint32), ('epoch', c_int64)]

AVCaptureDevice = ObjCClass('AVCaptureDevice')

_device = AVCaptureDevice.defaultDeviceWithDeviceType_mediaType_position_(
  'AVCaptureDeviceTypeBuiltInWideAngleCamera', 'vide', 1)

dur = _device.activeVideoMinFrameDuration(restype=CMTime, argtypes=[])
dur.value *= 3
_device.lockForConfiguration_(None)
_device.setActiveVideoMinFrameDuration_(dur, restype=None, argtypes=[CMTime])
_device.unlockForConfiguration()

@JonB 350 passes and still works. I'll play with parameters to see how it will go.

Thanks.