How to use stero calibration files

erik · Aug 15, 2022

Hi mrvladimir ,
Could you elaborate on the creating a stereo depth node in python? Are you using depthai's stereoDepth node? If that's the case, you don't need to do anything - calibration data is written on the OAK device and StereoDepth node will read from it to set up the stereo pipeline. Thoughts?
Thanks, Erik

theDesertMoon · Aug 15, 2022

As a side question on this subject. I have a few OAK cameras and I'm comparing their ability to do object tracking for speed measurements. The OAK-1-POE is a champ at this and is clocking in 30+ fps. The OAK-D-Lite is clocking in around 6 fps but is exhibiting after images. I see that the non-modular OAK devices aren't supposed to need recalibration.

The other thing that isn't evident in a still image is the after images just gradually move off screen not really following the bottle. The bottle doesn't travel through the rails but the objID's will just pick a direction and continue to move.

In regards to the 6 FPS is this normal? They both push _30fps using the same model and only the Color camera on the OAK-D. When using Spatial however I get a message compiling for 5 shaves would be better. Could that make that large of a difference? If so, that's cool as I need to rebuild my model anyway with some new data.

Mmrvladimir · Aug 15, 2022

Hi Eric. Yes, I'm using the StereoNode as in the docs: https://docs.luxonis.com/projects/api/en/latest/components/nodes/stereo_depth/.

Are the meshes also written to the device, or just the camera matrix plus distortion coefficients?

Those values are dependent on the input image resolution. When I decrease the left/right mono resolution, how does the depth node adjust the distortion mesh?

erik · Aug 18, 2022

Hi theDesertMoon ,
I apologize for late reply. 6 FPS definately isn't normal, and I assume it's an issue with either the transfer speed, or maybe syncing speed. COuld you share the code you are using? If it's running the same model, both OAK-D_Lite and OAK-1-POE should have the exact same performance.
Thanks, Erik

erik · Aug 18, 2022

Hi mrvladimir , they scale based on the resolution that you use, see here.
Thanks, Eirk

theDesertMoon · Aug 19, 2022

erik No need for apologies. I see these forums and your collective expertise is in high demand.

So both of these screenshots are from the OAK-D-Lite. The 'flat tracker' image uses a modified 'object_detection.py' script while 'spatial tracker' uses a modified 'spatial_object_detection.py' script.

They use similar models too. The main difference is that I ran the openvino conversion script twice - once for 6 shaves, used with 'o_d.py' and then I converted for 5 shaves, used with 's_o_d.py'. Changing the model almost doubled my framerate from 6 to 12 but the OAK-D-Lite's running 'o_d.py' still acheives 30 fps.

from pathlib import Path
import cv2
import depthai as dai
import numpy as np
import time
import argparse
import os

fullFrameTracking = True

CWD_PATH = os.getcwd()
MODEL_NAME = "bottles_tflite0"
GRAPH_NAME = "prodlibv007_6shave.blob"
nnPath = os.path.join(CWD_PATH,MODEL_NAME,GRAPH_NAME)

labelMap = ["background", "dbaseblu0dwn", "dbaseblu0top", "petalgrn0dwn", "petalgrn0top", "petalred0dwn", "petalred0top"]

pipeline = dai.Pipeline()

camRgb = pipeline.create(dai.node.ColorCamera)
detectionNetwork = pipeline.create(dai.node.MobileNetDetectionNetwork)
objectTracker = pipeline.create(dai.node.ObjectTracker)

xlinkOut = pipeline.create(dai.node.XLinkOut)
trackerOut = pipeline.create(dai.node.XLinkOut)

xlinkOut.setStreamName("preview")
trackerOut.setStreamName("tracklets")

camRgb.setPreviewSize(300, 300)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setIspScale(1,3)
camRgb.setInterleaved(False)
camRgb.setPreviewKeepAspectRatio(False)

detectionNetwork.setBlobPath(nnPath)
detectionNetwork.setConfidenceThreshold(0.1)
detectionNetwork.input.setBlocking(False)

objectTracker.setTrackerType(dai.TrackerType.ZERO_TERM_COLOR_HISTOGRAM)

objectTracker.setTrackerIdAssignmentPolicy(dai.TrackerIdAssignmentPolicy.SMALLEST_ID)

camRgb.preview.link(detectionNetwork.input)
objectTracker.passthroughTrackerFrame.link(xlinkOut.input)

if fullFrameTracking:
  camRgb.video.link(objectTracker.inputTrackerFrame)
else:
  detectionNetwork.passthrough.link(objectTracker.inputTrackerFrame)

detectionNetwork.passthrough.link(objectTracker.inputDetectionFrame)
detectionNetwork.out.link(objectTracker.inputDetections)
objectTracker.out.link(trackerOut.input)

found, device_info = dai.Device.getDeviceByMxId("18443010C1E8C11200") #OAK-D-Lite @ Pi

with dai.Device(pipeline, device_info) as device:

  preview = device.getOutputQueue("preview", 2, False)
  tracklets = device.getOutputQueue("tracklets", 2, False)

  startTime = time.monotonic()
  counter = 0
  fps = 0
  totalspeed = 0
  avragspeed = 0
  speedtimer = time.perf_counter()
  frame = None
  prevlock = []
  centers = []
  for i in range(100):
    prevlock.append([-1,-1,0])
  while(True):
    imgFrame = preview.get()
    track = tracklets.get()

    counter+=1
    current_time = time.monotonic()
    if (current_time - startTime) > 1 :
      fps = counter / (current_time - startTime)
      counter = 0
      startTime = current_time

    color = (255, 0, 0)
    frame = imgFrame.getCvFrame()
    trackletsData = track.tracklets
    tempoldvals = ""
    for t in trackletsData:
      roi = t.roi.denormalize(frame.shape[1], frame.shape[0])
      x1 = int(roi.topLeft().x)
      y1 = int(roi.topLeft().y)
      x2 = int(roi.bottomRight().x)
      y2 = int(roi.bottomRight().y)
      centerpoint = ((x2-x1)/2,(y2-y1)/2)
      if prevlock[t.id][1] == -1:
        prevlock[t.id][0] = y1
      else:
        prevlock[t.id][2] = prevlock[t.id][1]-y1
        prevlock[t.id][0] = y1
      try:
        label = labelMap[t.label]
      except:
        label = t.label
      totalspeed += prevlock[t.id][2]
      cv2.putText(frame, f"{[t.id]}", (x1 + 10, y1 + 35), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (250,250,250))
      cv2.rectangle(frame, (x1, y1), (x2, y2), color, cv2.FONT_HERSHEY_SIMPLEX)
    longrun = len(trackletsData)
    if longrun > 0:
      avragspeed += totalspeed/longrun #total speed
    temptimer = time.perf_counter() - speedtimer
    if temptimer > 1:
      if (avragspeed < 20):#2.5):
        avragspeed = 0
      totalspeed = 0
      avragspeed = 0
      speedtimer = time.perf_counter()
    for i in range(100):
      if prevlock[i][0] == -1:
        prevlock[i][1] = -1
        prevlock[i][2] = 0
      else:
        prevlock[i][1] = prevlock[i][0]
        prevlock[i][0] = -1
    
    cv2.putText(frame, "NN fps: {:.2f}".format(fps), (2, frame.shape[0] - 4), cv2.FONT_HERSHEY_TRIPLEX, 0.6, color)

    cv2.imshow("flat tracker", frame)

   if cv2.waitKey(1) == ord('q'):
      break

from pathlib import Path
import cv2
import depthai as dai
import numpy as np
import time
import argparse
import os

fullFrameTracking = True

CWD_PATH = os.getcwd()
MODEL_NAME = "bottles_tflite0"
GRAPH_NAME = "prodlibv007_5shave.blob"
nnPath = os.path.join(CWD_PATH,MODEL_NAME,GRAPH_NAME)

labelMap = ["background", "dbaseblu0dwn", "dbaseblu0top", "petalgrn0dwn", "petalgrn0top", "petalred0dwn", "petalred0top"]

pipeline = dai.Pipeline()

camRgb = pipeline.create(dai.node.ColorCamera)
spatialDetectionNetwork = pipeline.create(dai.node.MobileNetSpatialDetectionNetwork)
monoLeft = pipeline.create(dai.node.MonoCamera)
monoRight = pipeline.create(dai.node.MonoCamera)
stereo = pipeline.create(dai.node.StereoDepth)
objectTracker = pipeline.create(dai.node.ObjectTracker)

xoutRgb = pipeline.create(dai.node.XLinkOut)
trackerOut = pipeline.create(dai.node.XLinkOut)

xoutRgb.setStreamName("preview")
trackerOut.setStreamName("tracklets")

camRgb.setPreviewSize(300, 300)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setIspScale(1,3)
camRgb.setInterleaved(False)
camRgb.setPreviewKeepAspectRatio(False)

monoLeft.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
monoLeft.setBoardSocket(dai.CameraBoardSocket.LEFT)
monoRight.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
monoRight.setBoardSocket(dai.CameraBoardSocket.RIGHT)

stereo.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)

stereo.setDepthAlign(dai.CameraBoardSocket.RGB)

spatialDetectionNetwork.setBlobPath(nnPath)
spatialDetectionNetwork.setConfidenceThreshold(0.1)
spatialDetectionNetwork.input.setBlocking(False)
spatialDetectionNetwork.setBoundingBoxScaleFactor(0.5)
spatialDetectionNetwork.setDepthLowerThreshold(100)
spatialDetectionNetwork.setDepthUpperThreshold(5000)

objectTracker.setTrackerType(dai.TrackerType.ZERO_TERM_COLOR_HISTOGRAM)

objectTracker.setTrackerIdAssignmentPolicy(dai.TrackerIdAssignmentPolicy.SMALLEST_ID)

monoLeft.out.link(stereo.left)
monoRight.out.link(stereo.right)

camRgb.preview.link(spatialDetectionNetwork.input)
objectTracker.passthroughTrackerFrame.link(xoutRgb.input)
objectTracker.out.link(trackerOut.input)

if fullFrameTracking:
    camRgb.setPreviewKeepAspectRatio(False)
    camRgb.video.link(objectTracker.inputTrackerFrame)
    objectTracker.inputTrackerFrame.setBlocking(False)
    objectTracker.inputTrackerFrame.setQueueSize(2)
else:
    spatialDetectionNetwork.passthrough.link(objectTracker.inputTrackerFrame)

spatialDetectionNetwork.passthrough.link(objectTracker.inputDetectionFrame)
spatialDetectionNetwork.out.link(objectTracker.inputDetections)
stereo.depth.link(spatialDetectionNetwork.inputDepth)

found, device_info = dai.Device.getDeviceByMxId("18443010C1E8C11200") #OAK-D-Lite @ Pi

with dai.Device(pipeline, device_info) as device:

    preview = device.getOutputQueue("preview", 2, False)
    tracklets = device.getOutputQueue("tracklets", 2, False)

    startTime = time.monotonic()
    counter = 0
    fps = 0
    color = (255, 0, 0)

    while(True):
        imgFrame = preview.get()
        track = tracklets.get()

        counter+=1
        current_time = time.monotonic()
        if (current_time - startTime) > 1 :
            fps = counter / (current_time - startTime)
            counter = 0
            startTime = current_time

        frame = imgFrame.getCvFrame()
        trackletsData = track.tracklets
        for t in trackletsData:
            roi = t.roi.denormalize(frame.shape[1], frame.shape[0])
            x1 = int(roi.topLeft().x)
            y1 = int(roi.topLeft().y)
            x2 = int(roi.bottomRight().x)
            y2 = int(roi.bottomRight().y)
            centerpoint = ((x2-x1)/2,(y2-y1)/2)
            try:
                label = labelMap[t.label]
            except:
                label = t.label

            cv2.putText(frame, f"ID: {[t.id]}", (x1 + 10, y1 + 35), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (250,250,250))
            cv2.rectangle(frame, (x1, y1), (x2, y2), color, cv2.FONT_HERSHEY_SIMPLEX)

        cv2.putText(frame, "NN fps: {:.2f}".format(fps), (2, frame.shape[0] - 4), cv2.FONT_HERSHEY_TRIPLEX, 0.6, color)

        cv2.imshow("spatial tracker", frame)

        if cv2.waitKey(1) == ord('q'):
            break
`
--------------------------------------------------------------------------
--------------------------------------------------------------------------

erik · Aug 19, 2022

Hi theDesertMoon ,
I would be very interested in the models compiled for 5/6/12 shaves, so we can try & verify it locally, as we haven't seen such drastic performance improvements. By default, NN nodes (spatial / obj detector nodes as well) run 2 threads, so compiling for 12 isn't needed, and we recommend compiling for 1/2 of all available cores (so they can be distributed between the 2 threads).
If models are not considered public, you can also use my email: erik@luxonis.com

Thanks, Erik

theDesertMoon · Aug 22, 2022

erik Hi Erik! I think I just missed your reply before I left.

Here's a link to the two models. I've only compiled it for 6 shaves and 5 shaves (the depthai script recommends 5 shaves for the spatial_object_detection.py) however.

https://drive.google.com/drive/folders/1nXqFpaOOS7A92ITrjCkFlOE76jOUefLZ?usp=sharing

They aren't the best models but I'm working on that, too.

erik · Aug 22, 2022

Hi theDesertMoon ,
Could you maybe provide the MRE with everything zipped together, so it's easier to reproduce locally for us?
Thanks, Erik

theDesertMoon · Aug 22, 2022

erik Zipping I don't mind but is MRE 'mean relative error'? Unfortunately I didn't save any of the output from creating the model besides the frozen inference graph files. My browser updated and the page reloaded with all of the outputs wiped out.

erik · Aug 22, 2022

Sorry about that! I meant Minimal reproducible example, see how to create one here.
Thanks, Erik

theDesertMoon · Aug 23, 2022

erik Thank you, sir!

I think I've captured everything in the zip file here: https://drive.google.com/drive/folders/1nXqFpaOOS7A92ITrjCkFlOE76jOUefLZ?usp=sharing

I included additional notes in the readme file. Thanks for your input.

-Deon

erik · Aug 23, 2022

Hi theDesertMoon ,
There were a few issues with the MRE first (blobs were not where the code expected them to be, host side code was there (shouldn't be - should be minimal, a bunch of code commented out), but I managed to debug it. Here's the actual MRE.

In spatail_object_tracker.py you had this:

objectTracker.inputTrackerFrame.setQueueSize(2)

If you remove limiting this queue size to only 2 messages (4 by default iirc), it runs at +25FPS.
Thanks, Erik

theDesertMoon · Aug 23, 2022

erik Ah, so it was self-inflicted. Thank you. I'll clean up any future MRE's should I need to submit them.

Thank you very much.