1 of 41

2 of 41

Bias in Light-Based Sensors

Soap Dispenser

Chukwuemeka Afigbo ‏/ Twitter

3 of 41

Bias in Light-Based Sensors

Soap Dispenser

Pulse Oximeters

Infrared Thermometer

Facial Recognition

Plethysmography

Chukwuemeka Afigbo ‏/ Twitter

Cleveland Clinic, 2019

Grace Cary/Getty Images

Najibi, 2020

McDuff, 2020

4 of 41

Bias in Light-Based Sensors

Soap Dispenser

Chukwuemeka Afigbo ‏/ Twitter

5 of 41

Bias in Light-Based Sensors

Soap Dispenser

Pulse Oximeters

Infrared Thermometer

Facial Recognition

Plethysmography

Chukwuemeka Afigbo ‏/ Twitter

Cleveland Clinic, 2019

Grace Cary/Getty Images

Najibi, 2020

McDuff, 2020

6 of 41

Bias in Light-Based Sensors

Infrared Thermometer

Infrared Thermometer

Soap Dispenser

Pulse Oximeters

Plethysmography

7 of 41

Plethysmography

McDuff, 2015

8 of 41

Plethysmography

Captured Video

Clinical Finger Pulse Oximeter (GT)

Arbitrary units

Time

9 of 41

Plethysmography

Captured Video

Clinical Finger Pulse Oximeter (GT)

Arbitrary units

Time

10 of 41

Plethysmography

Captured Video

Pulse Oximeter Ground Truth - 96.0 BPM

Time

Arbitrary units

Cropped Face Frame

Waveform from Camera - 96.7 BPM

Time

Arbitrary units

ML

[Yu, 2019]

11 of 41

Plethysmography

Captured Video

Cropped Face Frame

Estimated (Blue) and GT (Green)

Time

Arbitrary units

12 of 41

Plethysmography

Captured Video

Cropped Face Frame

Estimated (Blue) and GT (Green)

Time

Arbitrary units

  • Heart Rate Detection

  • Oxygen Saturation
  • Blood Pressure Monitoring

  • Atrial Fibrillation Detection

13 of 41

Plethysmography with Darker Skin Tones

Estimated Waveform from Camera - 74.2

Arbitrary units

Arbitrary units

Pulse Oximeter Ground Truth - BPM 83.1

Arbitrary units

Time

Time

Cropped Face Frame

14 of 41

Plethysmography with Darker Skin Tones

Cropped Face Frame

Estimated (Blue) and GT (Green)

Time

Arbitrary units

15 of 41

Plethysmography with Darker Skin Tones

Cropped Face Frame

Estimated (Blue) and GT (Green)

Time

Arbitrary units

10% more darker skin tones fail cardiac monitoring standards* than lighter skin tones.

* Failure occurs when the percent error is greater than 10% from ground truth.

16 of 41

Contributions of This Work

  • Establishing bias of RGB camera and performance-

fairness tradeoff of RGB and radar

17 of 41

Contributions of This Work

  • Establishing bias of RGB camera and performance-

fairness tradeoff of RGB and radar

  • Propose a multimodal fusion between the

RGB camera and radar for improving performance

and fairness

18 of 41

Contributions of This Work

  • Establishing bias of RGB camera and performance-

fairness tradeoff of RGB and radar

  • Propose a multimodal fusion between the

RGB camera and radar for improving performance

and fairness

  • IRB Dataset with Diverse Skintone Representation

19 of 41

Radar-Plethysmography

0.25 mm�

This is how much your chest moves when the heart beats

https://bit.ly/3Ql8zDm

20 of 41

Radar-Plethysmography

https://bit.ly/3Ql8zDm

0.25 mm�

This is how much your chest moves when the heart beats

21 of 41

Radar-Plethysmography

Radar

Small Chest Vibrations

Radar (77 GHz)

has 10 µm range sensitivity

22 of 41

Radar-Plethysmography

Clinical Pulse Oximeter (GT)

Estimated from Radar

Time

Time

Arbitrary units

Arbitrary units

Radar Matrix

Time

Distance (Frequency)

23 of 41

Error-Bias Plot

1

2

1

2

0

24 of 41

Error-Bias Plot

HR Error (MAE, bpm)

1

2

1

2

0

25 of 41

Error-Bias Plot

HR Error (MAE, bpm)

1

2

1

2

0

Better

26 of 41

Error-Bias Plot

HR Error (MAE, bpm)

Bias (MAE, bpm)

1

2

1

2

0

Better

27 of 41

Error-Bias Plot

HR Error (MAE, bpm)

Bias (MAE, bpm)

1

2

1

2

0

Better

Better

28 of 41

Error-Bias Plot

HR Error (MAE, bpm)

Bias (MAE, bpm)

Radar

RGB

1

2

1

2

0

Better

Better

29 of 41

Fusing RGB and Radar

RGB Video

Radar Range Data

Fusion Architecture

time

Waveform

Ground truth

Estimated

PPG Loss

30 of 41

Beyond Bias in Plethysmography

HR Error (MAE, bpm)

Bias (MAE, bpm)

Radar

RGB

Our Naive

Fusion

1

2

1

2

0

Better

Better

31 of 41

Fusing RGB and Radar

RGB Video

Radar Range Data

Fusion Architecture

time

Waveform

Ground truth

Estimated

PPG Loss

Performance Loss!

32 of 41

Fusing RGB and Radar

RGB Video

Radar Range Data

Fusion Architecture

time

Waveform

Ground truth

Estimated

PPG Loss

Performance Loss!

Need to impose fairness constraint!

33 of 41

Fusing RGB and Radar

RGB Video

Radar Range Data

Fusion Architecture

time

Waveform

Ground truth

Estimated

PPG Loss

34 of 41

Error-Bias Plot

HR Error (MAE, bpm)

Bias (MAE, bpm)

Radar

RGB

Our Naive

Fusion

Fair Fusion

1

2

1

2

0

Better

Better

35 of 41

Error-Bias Plot

HR Error (MAE, bpm)

Bias (MAE, bpm)

Radar

RGB

Our Naive

Fusion

Fair Fusion

1

2

1

2

0

Better

Better

3% more dark skin subjects fail cardiac monitoring standard (compare with 10% for RGB)

36 of 41

Back to Darker Skin Tone Participant

Estimated (Purple) and GT (Green)

Time

Arbitrary units

Cropped Face Frame

Estimated HR = 81 BPM and GT HR = 83.1 BPM

37 of 41

Second Order Dataset Details and Results

91 Participants

Method

Performance (Fairness)

Fairness

MAE

MAPE

RMSE

PCC (r)

T-Test (APE %)

RGB [Yu 2019]

1.78 (2.22)

2.35% (2.63%)

5.26 (4.05)

0.91 (-0.25)

2.1, 12.2

Our RF

2.18 (0.51)

3.05% (0.69%)

6.12 (0.85)

0.89 (-0.13)

5.1, 8.4

Our Fair Fusion

1.12 (0.67)

1.52% (0.79%)

3.42 (1.44)

0.95 (-0.10)

1.1, 4.2

RGB Camera

Fusion

Radar

38 of 41

Towards Fair Computational Imaging

39 of 41

Towards Fair Computational Imaging

Pulse Oximeters

Thermometers

Facial Recognition

IR Sensors

Other devices

40 of 41

Towards Fair Computational Imaging

Other Axes of Bias

Body Mass

Error

Pulse Oximeters

Thermometers

Facial Recognition

IR Sensors

Other devices

41 of 41

Towards Fair Computational Imaging

Pulse Oximeters

Thermometers

Facial Recognition

IR Sensors

Other devices

Other Axes of Bias

Body Mass

Error

Completely Fair Devices

Error

Bias