No distortion coefficient and rectification matrix generated by calibration.py

ShivamSharma

Hi,

I have a setup of two OVO 9282 cameras with an Oak FFC 3P. I tried calibrating the camera with the calibration.py file, and the following were the results, which didn't include distortion coefficients. Also, on pressing esc at the end, no calibration results were saved in a file. My question is: why is the calibration result incomplete with no distortion coefficients and rectification matrix? Is there no distortion in these cameras? Why is this file not saving calibration data?

shivam157@ubuntu:~/depthai$ python3 calibrate.py -s 3.8 --board OAK-FFC-3P -nx 13 -ny 7 -dbg

Using dataset path: dataset

Starting image processing

<------------Calibrating right ------------>

INTRINSIC CALIBRATION

Reprojection error of right: 0.1450288774499517

<------------Calibrating left ------------>

INTRINSIC CALIBRATION

Reprojection error of left: 0.18671330931332783

<-------------Extrinsics calibration of right and left ------------>

Reprojection error is 0.1806323559925981

<-------------Epipolar error of right and left ------------>

Original intrinsics ....

L [[909.05184795 0. 644.25179619]

[ 0. 909.33572655 383.70089729]

[ 0. 0. 1. ]]

R: [[906.07261965 0. 625.2752514 ]

[ 0. 906.30579217 399.46458918]

[ 0. 0. 1. ]]

Intrinsics from the getOptimalNewCameraMatrix/Original ....

L: [[906.07261965 0. 625.2752514 ]

[ 0. 906.30579217 399.46458918]

[ 0. 0. 1. ]]

R: [[906.07261965 0. 625.2752514 ]

[ 0. 906.30579217 399.46458918]

[ 0. 0. 1. ]]

Average Epipolar Error is : 0.1324884918914444

Displaying Stereo Pair for visual inspection. Press the [ESC] key to exit.

Reprojection error threshold -> 1.1111111111111112

right Reprojection Error: 0.145029

/home/shivam157/depthai/calibrate.py:1029: DeprecationWarning: Conversion of an array with ndim > 0 to a scalar is deprecated, and will error in future. Ensure you extract a single element from your array before performing this operation. (Deprecated NumPy 1.25.)

calibration_handler.setDistortionCoefficients(stringToCam[camera], cam_info['dist_coeff'])

/home/shivam157/depthai/calibrate.py:1068: DeprecationWarning: Conversion of an array with ndim > 0 to a scalar is deprecated, and will error in future. Ensure you extract a single element from your array before performing this operation. (Deprecated NumPy 1.25.)

calibration_handler.setCameraExtrinsics(stringToCam[camera], stringToCam[cam_info['extrinsics']['to_cam']], cam_info['extrinsics']['rotation_matrix'], cam_info['extrinsics']['translation'], specTranslation)

Reprojection error threshold -> 1.1111111111111112

left Reprojection Error: 0.186713

[]

py: DONE.

shivam157@ubuntu:~/depthai$

```

The following is the board file:
{

"board_config":

{

    "name": "OAK-FFC-3P",

    "revision": "R1M0E1",

    "cameras":{

        "CAM_C": {

            "name": "right",

            "hfov":  71.86,

            "type": "mono",

            "extrinsics": {

                "to_cam": "CAM_B",

                "specTranslation": {

                    "x": 14.5,

                    "y": 0,

                    "z": 0

                },

                "rotation":{

                    "r": 0,

                    "p": 0,

                    "y": 0

                }

            }

        },

        "CAM_B": {

            "name": "left",

            "hfov":  71.86,

            "type": "mono"

            }

    },

    "stereo_config":{

        "left_cam": "CAM_B",

        "right_cam": "CAM_C"

    }

}

}

I think this is related to numpy version. I am not sure if I downgrade numpy that it will support Opencv 4.9. It looks like the error is in the distortion coefficient and extrinsic parameters lines of code. What should I do?

ShivamSharma

I am not 100% sure but the number of markers are different in some images. This might be the problem?

jakaskerl

Hi @ShivamSharma
And if you tried to load the calibration file with https://docs.luxonis.com/software/depthai/examples/calibration_reader/?

Thanks,
Jaka

ShivamSharma

When I run the script given in the link you sent, I get the following output:

venvshivam157@ubuntu:~/depthai-python$ python3 get_calib.py

RGB Camera Default intrinsics...

[[3133.269775390625, 0.0, 1922.1737060546875], [0.0, 3133.713623046875, 1102.6722412109375], [0.0, 0.0, 1.0]]

3840

2160

RGB Camera Default intrinsics...

[[3133.269775390625, 0.0, 1922.1737060546875], [0.0, 3133.713623046875, 1102.6722412109375], [0.0, 0.0, 1.0]]

3840

2160

RGB Camera resized intrinsics... 3840 x 2160

[[3.13326978e+03 0.00000000e+00 1.92217371e+03]

[0.00000000e+00 3.13371362e+03 1.10267224e+03]

[0.00000000e+00 0.00000000e+00 1.00000000e+00]]

RGB Camera resized intrinsics... 4056 x 3040

[[3.30951611e+03 0.00000000e+00 2.03029590e+03]

[0.00000000e+00 3.30998486e+03 1.54394751e+03]

[0.00000000e+00 0.00000000e+00 1.00000000e+00]]

LEFT Camera Default intrinsics...

[[693.9341430664062, 0.0, 664.17138671875], [0.0, 694.2032470703125, 406.8681640625], [0.0, 0.0, 1.0]]

1280

800

LEFT Camera resized intrinsics... 1280 x 720

[[693.93414307 0. 664.17138672]

[ 0. 694.20324707 366.86816406]

[ 0. 0. 1. ]]

RIGHT Camera resized intrinsics... 1280 x 720

[[690.59875488 0. 619.99499512]

[ 0. 690.54821777 276.21847534]

[ 0. 0. 1. ]]

LEFT Distortion Coefficients...

k1: -0.14615201950073242

k2: 5.131624221801758

p1: -0.0004437704337760806

p2: -0.0003282237739767879

k3: 0.816307544708252

k4: 0.22460998594760895

k5: 5.021973609924316

k6: 2.6677396297454834

s1: 0.0

s2: 0.0

s3: 0.0

s4: 0.0

τx: 0.0

τy: 0.0

RIGHT Distortion Coefficients...

k1: 1.1050113439559937

k2: 1.1275875568389893

p1: 0.00031773868249729276

p2: 0.0002416888455627486

k3: 0.12704245746135712

k4: 1.4751837253570557

k5: 1.46196711063385

k6: 0.46535491943359375

s1: 0.0

s2: 0.0

s3: 0.0

s4: 0.0

τx: 0.0

τy: 0.0

RGB FOV 68.7938003540039, Mono FOV 110.0

LEFT Camera stereo rectification matrix...

[[ 9.90124752e-01 7.40214346e-03 -3.64586040e+01]

[-7.32702106e-03 9.95746611e-01 -8.60322599e+01]

[-8.12898310e-06 3.72185724e-06 1.00401435e+00]]

RIGHT Camera stereo rectification matrix...

[[ 9.94906762e-01 7.44132256e-03 4.97689224e+00]

[-7.36240841e-03 1.00101704e+00 2.47433011e+00]

[-8.16824369e-06 3.74155680e-06 1.00401153e+00]]

Transformation matrix of where left Camera is W.R.T right Camera's optical center

[[ 9.99970078e-01 2.70099915e-03 7.25283753e-03 5.21282959e+00]

[-2.66335462e-03 9.99982953e-01 -5.19499322e-03 -1.26599353e-02]

[-7.26674590e-03 5.17552113e-03 9.99960184e-01 8.40604864e-03]

[ 0.00000000e+00 0.00000000e+00 0.00000000e+00 1.00000000e+00]]

Transformation matrix of where left Camera is W.R.T RGB Camera's optical center

[[ 9.99992728e-01 -2.86172348e-04 3.81197175e-03 7.94537354e+00]

[ 2.56973435e-04 9.99970675e-01 7.65808392e-03 -3.24345136e+00]

[-3.81405186e-03 -7.65704829e-03 9.99963403e-01 -5.16706049e-01]

[ 0.00000000e+00 0.00000000e+00 0.00000000e+00 1.00000000e+00]]

venvshivam157@ubuntu:~/depthai-python$

Is this calibration data from my camera? I am using two Mono OV9282 camera with each 75 degree FOV. I don't have any rgb camera.

I only used calibration.py file. I don't think I flashed the device with the current setup. My baseline is 14.5 cm in x

jakaskerl

Hi ShivamSharma
Looks like the flash didn't succeed.
What is -dbg parameter? Are you using the latest calibrate script? Also try disabling rgb with -dsb rgb.

Thanks,
Jaka

ShivamSharma

Thank you for your reply!

I am using the latest branch of the depthAI library. The dbg parameter is to make the calibration file use the dataset.

I will use the dsb tag with calibration file. Do you have more debugging steps?

jakaskerl

Hi @ShivamSharma

ShivamSharma The dbg parameter is to make the calibration file use the dataset.

Use -m process flag to only run the processing stage and not the image acquisition. The -dbg flag prevents the flashing part of the script.

Thanks,
Jaka

ShivamSharma

I used the -m process and -dsb rgb tags you mentioned and the calibration.py file flashed the device and the calibration result was saved in the resources folder. I will check how accurate it is by getting disparity and will let you know the result.

ShivamSharma

Does the following calibration result look correct to you? It does not look completely correct to me because I am using only two sockets on Oak FFC 3P and this data has 3 sockets 0,1, and 2. I don't know how it calculated rotation and translation between 2 and 0 because 2 doesn't exist. Is it safe to use the rotation, translation between 0 and 1? I will still have to calculate the projection matrix for both 0, and 1 from k, R and t.

Here is the result:
{

"batchName": "",

"batchTime": 1679712500,

"boardConf": "IR-C00M05-00",

"boardCustom": "",

"boardName": "DM1090",

"boardOptions": 0,

"boardRev": "R3M0E3",

"cameraData": [

[

2,

{

"cameraType": 0,

"distortionCoeff": [

1.1050113439559937,

1.1275875568389893,

0.00031773868249729276,

0.0002416888455627486,

0.12704245746135712,

1.4751837253570557,

1.46196711063385,

0.46535491943359375,

0.0,

0.0

],

"extrinsics": {

"rotationMatrix": [

[

0.9999896287918091,

-0.002969305729493499,

-0.0034563534427434206

],

[

0.003013428533449769,

0.9999131560325623,

0.012831280939280987

],

[

0.003417953150346875,

-0.012841563671827316,

0.9999117255210876

]

],

"specTranslation": {

"x": 2.5999999046325684,

"y": 3.299999952316284,

"z": -0.0

},

"toCameraSocket": 0,

"translation": {

"x": 2.7325892448425293,

"y": -3.2466089725494385,

"z": -0.5430911183357239

}

},

"height": 800,

"intrinsicMatrix": [

[

690.5987548828125,

0.0,

619.9949951171875

],

[

0.0,

690.5482177734375,

316.2184753417969

],

[

0.0,

1.0

]

],

"lensPosition": 0,

"specHfovDeg": 110.0,

"width": 1280

}

],

[

0,

{

"cameraType": 0,

"distortionCoeff": [

-2.396423578262329,

-10.477461814880371,

-0.0004928055568598211,

0.0011145062744617462,

26.935871124267578,

-2.4506843090057373,

-10.170122146606445,

26.494205474853516,

0.0,

0.0

],

"extrinsics": {

"rotationMatrix": [

[

0.997919499874115,

0.05123051628470421,

-0.03914157673716545

],

[

-0.051841288805007935,

0.9985463619232178,

-0.014751287177205086

],

[

0.03832896426320076,

0.016749747097492218,

0.9991247653961182

]

],

"specTranslation": {

"x": 14.5,

"y": 0.0,

"z": 0.0

},

"toCameraSocket": 1,

"translation": {

"x": -14.075519561767578,

"y": 0.5210347771644592,

"z": -0.29575979709625244

}

},

"height": 800,

"intrinsicMatrix": [

[

909.0518188476563,

0.0,

644.2517700195313

],

[

0.0,

909.3357543945313,

383.7008972167969

],

[

0.0,

1.0

]

],

"lensPosition": 135,

"specHfovDeg": 75.0,

"width": 1280

}

],

[

1,

{

"cameraType": 0,

"distortionCoeff": [

-3.9288971424102783,

-3.196507692337036,

-0.00022527640976477414,

0.0015712999738752842,

19.68087387084961,

-3.980436325073242,

-2.887248992919922,

19.199928283691406,

0.0,

0.0

],

"extrinsics": {

"rotationMatrix": [

[

0.9999700784683228,

0.0027009991463273764,

0.007252837531268597

],

[

-0.0026633546222001314,

0.9999829530715942,

-0.00519499322399497

],

[

-0.007266745902597904,

0.00517552113160491,

0.99996018409729

]

],

"specTranslation": {

"x": 5.300000190734863,

"y": -0.0,

"z": -0.0

},

"toCameraSocket": 2,

"translation": {

"x": 5.21282958984375,

"y": -0.012659935280680656,

"z": 0.008406048640608788

}

},

"height": 800,

"intrinsicMatrix": [

[

906.0726318359375,

0.0,

625.2752685546875

],

[

0.0,

906.3057861328125,

399.464599609375

],

[

0.0,

1.0

]

],

"lensPosition": 0,

"specHfovDeg": 75.0,

"width": 1280

}

]

],

"deviceName": "",

"hardwareConf": "F0-FV00-BC000",

"housingExtrinsics": {

"rotationMatrix": [],

"specTranslation": {

"x": 0.0,

"y": 0.0,

"z": 0.0

},

"toCameraSocket": -1,

"translation": {

"x": 0.0,

"y": 0.0,

"z": 0.0

}

},

"imuExtrinsics": {

"rotationMatrix": [

[

0.0,

0.0

],

[

0.0,

0.0

],

[

0.0,

0.0

]

],

"specTranslation": {

"x": 0.0,

"y": 0.0,

"z": 0.0

},

"toCameraSocket": -1,

"translation": {

"x": 0.0,

"y": 0.0,

"z": 0.0

}

},

"miscellaneousData": [],

"productName": "OAK-FFC-3P",

"stereoEnableDistortionCorrection": false,

"stereoRectificationData": {

"leftCameraSocket": 1,

"rectifiedRotationLeft": [

[

0.9990953207015991,

-0.03698360174894333,

0.020993344485759735

],

[

0.036815181374549866,

0.9992871880531311,

0.008353326469659805

],

[

-0.02128731645643711,

-0.007572895381599665,

0.999744713306427

]

],

"rectifiedRotationRight": [

[

0.9997386932373047,

0.014605958014726639,

-0.017585638910531998

],

[

-0.014745572581887245,

0.9998605847358704,

-0.007835760712623596

],

[

0.017468737438321114,

0.008093023672699928,

0.9998146295547485

]

],

"rightCameraSocket": 0

},

"stereoUseSpecTranslation": true,

"version": 7,

"verticalCameraSocket": -1

}

jakaskerl

ShivamSharma
Yes, they normally large due to distortion correction being a rational function and not only a polynomial one as seen here
https://docs.opencv.org/4.x/d9/d0c/group__calib3d.html

ShivamSharma

jakaskerl I think you didn’t see my updated reply

jakaskerl

ShivamSharma
Are you sure the correct board config is used? This should not happen, unless this is some other calibration file.

Thanks,
Jaka

ShivamSharma

It may be possible that I am providing incorrect camera name such as CAM_A and CAM_B in the board file. I will paste the board file I am using at the end. I am using the left and right socket of OAK FFC 3P. When I run ros2 launch depthai_ros_driver camera.launch.py it says -

[component_container-1] [14442C10E14FCBD600] [1.2.1] [2.336] [MonoCamera(3)] [error] Camera not detected on socket: 0

It is using the board file that I am passing to it because I can see HFOV and baseline is what I entered in the board file.

The board file is :

{
    "board_config": {
        "name": "OAK-FFC-3P",
        "revision": "R3M0E3",
        "cameras": {
            "CAM_A": {
                "name": "right",
                "hfov": 75,
                "type": "mono",
                "extrinsics": {
                    "to_cam": "CAM_B",
                    "specTranslation": {
                        "x": 14.5,
                        "y": 0,
                        "z": 0
                    },
                    "rotation": {
                        "r": 0,
                        "p": 0,
                        "y": 0
                    }
                }
            },
            "CAM_B": {
                "name": "left",
                "hfov": 75,
                "type": "mono"
            }
        },
        "stereo_config": {
            "left_cam": "CAM_B",
            "right_cam": "CAM_A"
        }
    }
}

The command is :
python3 calibrate.py -s 3.8 -brd OAK-FFC-3P.json -nx 13 -ny 7 -m process -dsb rgb

jakaskerl

Hi @ShivamSharma
If you are running with ROS, perhaps the config you are using is wrong? You need to specify the params_file and pass in a config file that doesn't use the RGB camera.
Examples: luxonis/depthai-rostree/humble/depthai_ros_driver/config

If you want to test the accuracy of depth, I'd suggest using some raw-depthai examples from depthai-python or depthai-core repository.

Thanks,
Jaka

ShivamSharma

When I used depthai python to see the disparity, the results were bad. I flashed the device with the board file in my reply above. I still have question about why the calibration file saved has 3 sockets when I am using -dsb rgb and I am using two mono camera with Oak FFC 3P. I will paste my command and result below:
python3 calibrate.py -s 3.8 -brd OAK-FFC-3P.json -nx 13 -ny 7 -m process -dsb rgb
python3 depth_preview_sr.py

Can you please help me understand why the disparity is so bad?

The following is the settings for the depth_preview file:

#!/usr/bin/env python3

import cv2
import depthai as dai
import numpy as np

# Closer-in minimum depth, disparity range is doubled (from 95 to 190):
extendedDisparity = False
# Better accuracy for longer distance, fractional disparity 32-levels:
subpixel = True
# Better handling for occlusions:
lrCheck = True

enableRectified = True

# Create pipeline
pipeline = dai.Pipeline()

# Define sources and outputs
left = pipeline.create(dai.node.ColorCamera)
right = pipeline.create(dai.node.ColorCamera)

# Create stereo
stereo = pipeline.create(dai.node.StereoDepth)
xoutDepth = pipeline.create(dai.node.XLinkOut)
xoutDepth.setStreamName("disparity")

# Properties
left.setResolution(dai.ColorCameraProperties.SensorResolution.THE_800_P)
left.setCamera("left")

right.setResolution(dai.ColorCameraProperties.SensorResolution.THE_800_P)
right.setCamera("right")

stereo.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)
stereo.initialConfig.setMedianFilter(dai.MedianFilter.KERNEL_7x7)
stereo.setLeftRightCheck(lrCheck)
stereo.setExtendedDisparity(extendedDisparity)
stereo.setSubpixel(subpixel)


# Linking
left.isp.link(stereo.left)
right.isp.link(stereo.right)
if enableRectified:
    xoutRectR = pipeline.create(dai.node.XLinkOut)
    xoutRectL = pipeline.create(dai.node.XLinkOut)
    xoutRectR.setStreamName("rectifiedRight")
    xoutRectL.setStreamName("rectifiedLeft")
    stereo.rectifiedLeft.link(xoutRectL.input)
    stereo.rectifiedRight.link(xoutRectR.input)
stereo.disparity.link(xoutDepth.input)

maxDisp = stereo.initialConfig.getMaxDisparity()

# Connect to device and start pipeline
with dai.Device(pipeline) as device:
    while not device.isClosed():
        queueNames = device.getQueueEvents()
        for q in queueNames:
            message = device.getOutputQueue(q).get()
            # Display arrived frames
            if type(message) == dai.ImgFrame:
                frame = message.getCvFrame()
                if 'disparity' in q:
                    disp = (frame * (255.0 / maxDisp)).astype(np.uint8)
                    disp = cv2.applyColorMap(disp, cv2.COLORMAP_JET)
                    cv2.imshow(q, disp)
                else:
                    cv2.imshow(q, frame)
        if cv2.waitKey(1) == ord('q'):
            break

jakaskerl

Hi @ShivamSharma
Yes, that is a bad calibration.

Few things to try:

swap left and right sockets, then view the preview again, hopefully only the sockets are swapped in calibration:

right.setCamera("left")
left.setCamera("right")

You are using OV9282 which is basically the same as OV9782, so the 3P can sometimes confuse them. Edit the calibration json to manually set the sensor model:

"CAM_B": {
                "model": "OV9282",
                "name": "left",
                ....
                "type": "mono",

for both. Make sure that you are running on the updated develop branch of depthai repo.

I see you are using an altered depth preview script where the camera nodes are ColorCamera. How come? 9282 is mono.

Thanks,
Jaka

ShivamSharma

Thanks!

I changed the sockets as you suggested and now the following are the results which look better as compared to before but I am not sure how accurate should I expect it to be.

I am using the depth_preview.py file. I can see the curtains in the disparity and the bed but I am also sitting in front of the camera and it is showing me as black.

Can the performance be improved further? Also, should I switch the sockets when I run the ros2 driver or I can just switch the calib file for each camera?

I was using the OV9282 model name in the board file before.

jakaskerl

Hi @ShivamSharma
14.5cm baseline is quite large, are you sure you are not sitting too close to the cameras?

ShivamSharma Can the performance be improved further? Also, should I switch the sockets when I run the ros2 driver or I can just switch the calib file for each camera?

A good recalibration is the only way to improve the depth. You can alter the calibration file (with the examples we have for doing it) and just swap the sockets.

Thanks,
Jaka

ShivamSharma

Yes, I am sitting close to the camera. I am still not getting good depth in ROS 2 from the cameras. I swapped the sockets in ROS2.

jakaskerl

ShivamSharma
Could you send over the dataset you have used (should be stored in the dataset folder)? Also left and right (not rectified) image for scene where the depth doesn't look ok, and a calibration dump. We'll try to reproduce the results to see what is wrong.

Thanks,
Jaka