Skip to content

Commit 8789f70

Browse files
committed
first version of model cards for bodyPose, bodySegmentation, handMesh, faceMesh
1 parent 1c0ee61 commit 8789f70

File tree

14 files changed

+269
-59
lines changed

14 files changed

+269
-59
lines changed

content/blog/cosa-nyu-ml-tools/index.md

+1-1
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,5 @@
11
---
2-
templateKey: #blog-post
2+
templateKey: blog-post
33
title: COSA x NYU Machine Learning Tools for Creative Coding
44
author: ml5.js
55
description: Join us at ITP for an informal series of talks and workshops exploring open-source machine learning tools for creative coding, presented in partnership with the Clinic for Open Source Arts (COSA)!
546 KB
Loading
13.9 KB
Loading

content/model-card/bodypose/index.md

+69
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,69 @@
1+
---
2+
templateKey: blog-post # <-- Uncomment this so that this post can be included in the blog list
3+
title: BodyPose Model Card
4+
author: ml5.js
5+
description: Friendly Machine Learning for the Web
6+
keywords: bias, model card, BodyPose
7+
image: "./images/_thumb.jpg"
8+
externalLink: (link)
9+
date: "2025-03-14" # YYYY-MM-DD or YYYY-MM-DD to YYYY-MM-DD or YYYY-MM-DD, YYYY-MM-DD, YYYY-MM-DD
10+
tags:
11+
- BodyPose
12+
featured: true
13+
---
14+
BodyPose is developed leveraging TensorFlow's [MoveNet](https://www.tensorflow.org/hub/tutorials/movenet#:~:text=MoveNet%20is%20an%20ultra%20fast,known%20as%20Lightning%20and%20Thunder) and [BlazePose](https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker) models.
15+
16+
______
17+
## MoveNet
18+
MoveNet was trained on [two datasets](https://storage.googleapis.com/movenet/MoveNet.SinglePose%20Model%20Card.pdf):
19+
20+
**COCO Keypoint Dataset Training Set 2017**
21+
- Date created: **2017**
22+
- Size: **28K images**
23+
- How the data was collected: “In-the-wild images with diverse scenes, instance sizes, and occlusions.” The original dataset of 64K images was distilled to the final 28K to only include three or less people per image.
24+
- Bias:
25+
* According to the public [model card](https://storage.googleapis.com/movenet/MoveNet.SinglePose%20Model%20Card.pdf), the qualitative analysis shows that although the dataset has a 3:1 male to female ratio, favors young and light skinned individuals, the models is stated to perform “fairly” (< 5% performance differences between most categories).
26+
* Categories of evaluation:
27+
* Male / Female (gender)
28+
* Young / Middle-age / Old (age)
29+
* Darker / Medium/ Lighter (skin tone)
30+
* There has been a fair amount of [research](https://medium.com/@rxtang/mitigating-gender-bias-in-captioning-systems-5a956e1e0d6d#:~:text=COCO%20dataset%20has%20an%20imbalanced,the%20bias%20learned%20by%20models) about the COCO Dataset. Most show that the dataset has numerous biases occurring due to underrepresentation of certain demographics.
31+
32+
**Active Dataset Training Set**
33+
- Date created: **2017-2021** ([assuming](https://blog.tensorflow.org/2021/05/next-generation-pose-detection-with-movenet-and-tensorflowjs.html))
34+
- Size: **23.5k images**
35+
- How the data was collected: “Images sampled from **YouTube fitness videos** which capture people exercising (e.g. HIIT, weight-lifting, etc.), stretching, or dancing. It contains diverse poses and motion with more motion blur and self-occlusions.”
36+
- Bias:
37+
* According to the model card, the models are stated to perform “fairly” (< 5% performance differences between all categories).
38+
* Categories of evaluation:
39+
* Male / Female (gender)
40+
* Young / Middle-age / Old (age)
41+
* Darker / Medium/ Lighter (skin tone)
42+
* The Active Single Person Image set, unlike COCO dataset, is not public, hence there is no additional research conducted to evaluate the fairness.
43+
44+
As stated, fitness videos uploaded to YouTube were used to assemble this internal Google dataset. Only in [2024](https://support.google.com/youtube/thread/313644973/third-party-ai-trainability-on-youtube?hl=en), Google [has provided](https://support.google.com/youtube/answer/15509945?hl=en) creators the opportunity to opt-out from Google using their videos for their AI/ML research.
45+
46+
___
47+
## BlazePose
48+
BlazePose’s [research paper](https://arxiv.org/pdf/2006.10204) and [model card](https://drive.google.com/file/d/10WlcTvrQnR_R2TdTmKw0nkyRLqrwNkWU/preview)
49+
- Date created: **2020-2021 (assuming)**
50+
- Size: **80K**
51+
- How the data was collected: Not stated in the original research paper. The model card asserts: “This model was trained and evaluated on images, including consented images (30K), of people using a mobile AR application captured with smartphone cameras in various “in-the-wild” conditions. The majority of training images (85K) capture a wide range of fitness poses.”
52+
- Bias:
53+
* According to the model card, the models are stated to perform “fairly”.
54+
* Categories of evaluation:
55+
* 14 subregions
56+
* Male / Female (gender)
57+
* 6 skin tones
58+
* Evaluation results:
59+
* Subregions (14): difference in confidence between average and worst performing regions of 4.8% for the heavy, 4.8% for the full and 6.5% for the light model.
60+
* Gender: difference in confidence is 1.1% for the heavy model, 2.2% for the full model and 3.1% for the lite model.
61+
* Skin tones: difference in confidence between worst and best performing categories is 5.7% for the heavy model, 7.0% for the full model and 7.3% for the lite model.
62+
63+
There is no additional research conducted to evaluate the fairness.
64+
There is no specific information on how the **consent** was obtained to get the images.
65+
66+
67+
____
68+
69+
#### Please submit any feedback/information you belive would be useful regarding this model [here](https://forms.gle/BPG44g3cJywSKjde6).
Loading
Loading
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
1+
---
2+
templateKey: blog-post # <-- Uncomment this so that this post can be included in the blog list
3+
title: BodySegmentation Model Card
4+
author: ml5.js
5+
description: Friendly Machine Learning for the Web
6+
keywords: bias, model card, BodySegmentation
7+
image: "./images/_thumb.jpg"
8+
externalLink: (link)
9+
date: "2025-03-14" # YYYY-MM-DD or YYYY-MM-DD to YYYY-MM-DD or YYYY-MM-DD, YYYY-MM-DD, YYYY-MM-DD
10+
tags:
11+
- BodySegmentation
12+
featured: true
13+
---
14+
The ml5.js BodySegmentation provides two models, **SelfieSegmentation** and **BodyPix**:
15+
16+
______
17+
## SelfieSegmentation
18+
19+
**MediaPipe Selfie Segmentation [Model Card](https://storage.googleapis.com/mediapipe-assets/Model%20Card%20MediaPipe%20Selfie%20Segmentation.pdf)**
20+
- Date created: **2021**
21+
- Size: **Not stated**
22+
- How the data was collected: “This model was trained and evaluated on images, including consented images of people using a mobile AR application captured with smartphone cameras in various “in-the-wild” conditions.”
23+
- Bias:
24+
* Categories of evaluation:
25+
* 17 demographical subregions
26+
* 6 skin tones
27+
* Male / Female (gender)
28+
* Evalutation results:
29+
* Subregions: Difference in confidence between average and worst performing regions of 1.11% for the general model, and 1.28% for the landscape model, lower than the criteria.
30+
* Gender: Differences in confidence are 1.6% for the general model and 0.6% for the landscape model.
31+
* Skin tone: Difference in confidence between worst and best performing
32+
33+
____
34+
## BodyPix
35+
This [short article](https://medium.com/tensorflow/introducing-bodypix-real-time-person-segmentation-in-the-browser-with-tensorflow-js-f1948126c2a0) is the only information on BodyPix that we have found.
36+
37+
____
38+
39+
#### Please submit any feedback/information you belive would be useful regarding this model [here](https://forms.gle/BPG44g3cJywSKjde6).
546 KB
Loading
13.9 KB
Loading

content/model-card/facemesh/index.md

+63
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,63 @@
1+
---
2+
templateKey: blog-post # <-- Uncomment this so that this post can be included in the blog list
3+
title: Facemesh Model Card
4+
author: ml5.js
5+
description: Friendly Machine Learning for the Web
6+
keywords: bias, model card, HandPose
7+
image: "./images/_thumb.jpg"
8+
externalLink: (link)
9+
date: "2025-03-14" # YYYY-MM-DD or YYYY-MM-DD to YYYY-MM-DD or YYYY-MM-DD, YYYY-MM-DD, YYYY-MM-DD
10+
tags:
11+
- HandPose
12+
featured: true
13+
---
14+
15+
## MediaPipe Face Mesh
16+
17+
**These are the [Model Card](https://drive.google.com/file/d/1sv4sSb9BSNVZhLzxXJ0jBv9DqD-4jnAz/view?pli=1), the [Research](https://arxiv.org/pdf/1907.06724), and the [Research Blog](https://sites.google.com/view/perception-cv4arvr/facemesh)**
18+
- Date created: **2018**
19+
- Size: **Not stated**
20+
- How the data was collected: “All dataset images were captured on a diverse set of smartphone cameras, both front- and back-facing. All images were captured in a real-world environment with different light, noise and motion conditions via an AR (Augmented Reality) application.”
21+
22+
- Bias:
23+
* According to the model card, the models are stated to perform “well” across most groups.
24+
* Categories of evaluation:
25+
* 17 geographic subregions
26+
* 6 skin tones
27+
* Male / Female (gender)
28+
* Evaluation results:
29+
* Subregions: Difference in confidence between best and worst performing regions of 0.9% for the tracking mode and 1.56% for the reacquisition mode.
30+
* Genders: Difference in confidence is 0.02% for the tracking mode and 0.1% for the reacquisition mode.
31+
* Skin tones: Difference in confidence is 0.24% for tracking mode and 1.12% for the reacquisition mode.
32+
* There is no additional research conducted to evaluate the fairness.
33+
34+
35+
36+
* There is no additional research conducted to evaluate the fairness.
37+
38+
_____
39+
## MediaPipe Attention Mesh
40+
**This is the [Model Card](https://drive.google.com/file/d/1tV7EJb3XgMS7FwOErTgLU1ZocYyNmwlf/preview)**
41+
- Date created: **2020**
42+
- Size: **30K (assuming)**
43+
- How the data was collected: “All dataset images were captured on a diverse set of smartphone cameras, both front- and back-facing. All images were captured in a real-world environment with different light, noise and motion conditions via an AR (Augmented Reality) application.”
44+
45+
- Bias:
46+
* According to the model card, the models are stated to perform “well” across most groups.
47+
48+
* Categories of evaluation:
49+
* 17 geographic subregions
50+
* 6 skin tones
51+
* Male / Female (gender)
52+
* Evaluation results:
53+
* Subregions: Difference in confidence between best and worst performing regions of 1.22% for the tracking mode and 1.27% for the reacquisition mode.
54+
* Gender: Difference in confidence is 0.01% for the tracking mode and 0.03% for the reacquisition mode.
55+
* Skin tones: Difference in confidence is 0.54% for tracking mode and 0.88% for the reacquisition mode.
56+
57+
- Potential bias:
58+
* Potential biases in the model which may result in drastic confidence differences if evaluated: May have difficulty with facial accessories like glasses or cultural headwear.
59+
60+
____
61+
62+
#### Please submit any feedback/information you belive would be useful regarding this model [here](https://forms.gle/BPG44g3cJywSKjde6).
63+
546 KB
Loading
13.9 KB
Loading

content/model-card/handpose/index.md

+29
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
---
2+
templateKey: blog-post # <-- Uncomment this so that this post can be included in the blog list
3+
title: HandPose Model Card
4+
author: ml5.js
5+
description: Friendly Machine Learning for the Web
6+
keywords: bias, model card, HandPose
7+
image: "./images/_thumb.jpg"
8+
externalLink: (link)
9+
date: "2025-03-14" # YYYY-MM-DD or YYYY-MM-DD to YYYY-MM-DD or YYYY-MM-DD, YYYY-MM-DD, YYYY-MM-DD
10+
tags:
11+
- HandPose
12+
featured: true
13+
---
14+
## Hand Detection / Tracking
15+
16+
**Hand Detection/Tracking [Model Card](https://drive.google.com/file/d/1sv4sSb9BSNVZhLzxXJ0jBv9DqD-4jnAz/view?pli=1)**
17+
- Date created: **2021** (assuming)
18+
- Size: **Not stated**
19+
- How the data was collected: “This model was trained and evaluated on images of people using a mobile AR application captured with smartphone cameras in various “in-the-wild” conditions.”
20+
21+
- Bias:
22+
* No evaluation has been conducted on this model. No access to the dataset was provided.
23+
* As stated on the model card: “as with many human sensing tools, performance may vary across skin tones, gender, age, and potentially other sensitive demographic characteristics.”
24+
* There is no additional research conducted to evaluate the fairness.
25+
26+
____
27+
28+
#### Please submit any feedback/information you belive would be useful regarding this model [here](https://forms.gle/BPG44g3cJywSKjde6).
29+

0 commit comments

Comments
 (0)