Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
© 2021 BASF
Object Detection and Dataset
Labeling Using Colors of
Manufactured Objects
Ian Childers, PhD
Head of Technolog...
© 2021 BASF
Coatings – We are the Leader In Color Design
2
Global color collection Annual trend book
Global design support...
© 2021 BASF
Object Detection with CNNs
• Convolutional neural networks learn
features and how the presence and
arrangement...
© 2021 BASF
Object Detection Edge Cases
Ultimate problem is the lack of a universal ground truth
Partial
occlusions
Flexib...
© 2021 BASF
Color and “Ground Truth”
• Color seems like an obvious choice for a
ground truth candidate
• Unfortunately, an...
© 2021 BASF
Fluorescent Chroma
• Fluorescence is the absorption of light at one
energy and emission of light at another en...
© 2021 BASF
Examples of Fluorescent Objects
7
Similar colors, different color IDs
3D shapes
• Commercial articles
• Plasti...
© 2021 BASF
System Design
8
Cameras
LEDs
Bandpass
Filters
System designed in
partnership with
Basic idea: block reflective...
© 2021 BASF
System Operation
9
Capture speed Classification demo
© 2021 BASF
Classification Model and Accuracy
• Can use many models for classification,
but nearest neighbors with
neighbo...
© 2021 BASF
Use Cases and Partnerships
11
Collect
images
Determine
labelling
criteria
Manually
label
images
Train
CNN
Depl...
© 2021 BASF
Conclusions
• Fluorescent chroma can be separated
from reflectance
• Fluorescent chroma can be used to
improve...
© 2021 BASF
Resource Slide
13
2021 Embedded Vision Summit
Color based Object Detection System
for Visual AI Applications D...
Upcoming SlideShare
Loading in …5
×

of

“Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 1 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 2 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 3 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 4 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 5 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 6 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 7 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 8 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 9 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 10 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 11 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 12 “Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF Slide 13
Upcoming SlideShare
What to Upload to SlideShare
Next
Download to read offline and view in fullscreen.

0 Likes

Share

Download to read offline

“Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF

Download to read offline

For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2021/07/object-detection-and-dataset-labeling-using-colors-of-manufactured-objects-a-presentation-from-basf/

Ian Childers, Head of Technology for Functional Coatings—Object Recognition at BASF, presents the “Object Detection and Dataset Labeling Using Colors of Manufactured Objects” tutorial at the May 2021 Embedded Vision Summit.

This talk introduces a new method for object detection for consumer goods and other applications based on measuring an object’s illumination invariant fluorescent chroma. Chroma is a measure of the colorfulness of an object relative to a similarly illuminated white object. Reflective chroma changes with the illumination, making it a poor choice for object detection in different environments. Fluorescent chroma, however, is nearly illumination invariant, making it an ideal input for object detection algorithms if the fluoresced light can be separated from the reflected light.

Childers describes simple designs of illuminators, cameras, and filters to achieve that separation, and shows the classification accuracy of the system to be 95-100% under different lighting conditions using the nearest neighbors algorithm with neighborhood component analysis. BASF has designed hundreds of unique colors and coatings for consumer goods and other items that can be used in this application and the system has many advantages compared to CNN object detection systems, as it does not require extensive training sets or an unoccluded view of the object.

Related Books

Free with a 30 day trial from Scribd

See all

Related Audiobooks

Free with a 30 day trial from Scribd

See all
  • Be the first to like this

“Object Detection and Dataset Labeling Using Colors of Manufactured Objects,” a Presentation from BASF

  1. 1. © 2021 BASF Object Detection and Dataset Labeling Using Colors of Manufactured Objects Ian Childers, PhD Head of Technology BASF Surface Solutions, Functional Coatings - Object Recognition
  2. 2. © 2021 BASF Coatings – We are the Leader In Color Design 2 Global color collection Annual trend book Global design support Innovative color concepts Key success factors: ◼ Trend research Excellent trend research and visualization ◼ Global color collection Translation of trend research into annual global color collection ◼ Innovative color concepts New inspiring color concepts with future- oriented approach ◼ Global design support Close to the customer all around the world
  3. 3. © 2021 BASF Object Detection with CNNs • Convolutional neural networks learn features and how the presence and arrangement of the features correlate to classes • Large numbers of items, almost infinite arrangements and views, different illumination • Lots of great work at solving these issues Remaining Problems 1. High quality, labelled training image sets are time consuming and expensive 2. AI predictions are less useful when the use case requires more certainty Figure from: Chen, Y.; Chen, R.; Liu, M.; Xiao, A.; Wu, D.; Zhao, S. Indoor Visual Positioning Aided by CNN-Based Image Retrieval: Training-Free, 3D Modeling- Free. Sensors 2018, 18, 2692. https://doi.org/10.3390/s18082692 3
  4. 4. © 2021 BASF Object Detection Edge Cases Ultimate problem is the lack of a universal ground truth Partial occlusions Flexible shapes Pose dependency Logos on other items Barcode issues Recycling Bin Top View Real life example 4
  5. 5. © 2021 BASF Color and “Ground Truth” • Color seems like an obvious choice for a ground truth candidate • Unfortunately, an object’s perceived color is strongly dependent upon illumination • We really want an easy measurement an object’s reflectance curve- it’s inherent colorfulness 5
  6. 6. © 2021 BASF Fluorescent Chroma • Fluorescence is the absorption of light at one energy and emission of light at another energy (ex. blue→red) • Magnitude of emission changes but shape (chroma) does not under different illumination • Ideal “ground truth” IF you can separate from reflectance • BASF has developed fluorescent coatings and camera/lighting designs to extract the fluorescence chroma and use it for object detection, classification, and segmentation 6
  7. 7. © 2021 BASF Examples of Fluorescent Objects 7 Similar colors, different color IDs 3D shapes • Commercial articles • Plastics • Printed packages • Fabrics • Printable labels • Post-production solution (similar to an added price sticker) • Custom coatings 7
  8. 8. © 2021 BASF System Design 8 Cameras LEDs Bandpass Filters System designed in partnership with Basic idea: block reflective light from reaching camera
  9. 9. © 2021 BASF System Operation 9 Capture speed Classification demo
  10. 10. © 2021 BASF Classification Model and Accuracy • Can use many models for classification, but nearest neighbors with neighborhood component analysis works well • Fluorescence segmentation is part of preprocessing • Best results with all channels, fluorescence doing majority of work • Some LEDs are better than others, both cameras are better than either one 10 # LEDs Camera # Channel Pixel Accuracy Accuracy Rank Object Accuracy 6 1,2 all 0.957 1 1 4 1,2 all 0.946 4 1 2 1,2 all 0.885 8 0.996 2 2 all 0.869 9 0.989 2 1 all 0.766 17 0.95 8 1,2 fluorescence 0.864 10 0.966 10 1,2 reflectance 0.62 20 0.765 Reflective Image Fluorescent Segmentation 280 classes, different lighting conditions for training and testing
  11. 11. © 2021 BASF Use Cases and Partnerships 11 Collect images Determine labelling criteria Manually label images Train CNN Deploy Update & retrain Deploy As is: Manual image labelling slows whole pipeline Color Detect articles Collect images Label with Color Detect Train CNN Deploy Update & retrain To be: Color Detect automatically labels images, increasing workflow rate and retraining speed Deploy Visual AI Training Retail Get in touch (colordetect@basf.com) if you think Color Detect technology could help your business or project! Packaged Goods Commerce
  12. 12. © 2021 BASF Conclusions • Fluorescent chroma can be separated from reflectance • Fluorescent chroma can be used to improve object classification and segmentation model performance • Applications in retail, visual AI model training, and more 12 Thank you!
  13. 13. © 2021 BASF Resource Slide 13 2021 Embedded Vision Summit Color based Object Detection System for Visual AI Applications Demo https://www.colordetect.basf.com https://www.edge-ai- vision.com/companies/basf/

For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2021/07/object-detection-and-dataset-labeling-using-colors-of-manufactured-objects-a-presentation-from-basf/ Ian Childers, Head of Technology for Functional Coatings—Object Recognition at BASF, presents the “Object Detection and Dataset Labeling Using Colors of Manufactured Objects” tutorial at the May 2021 Embedded Vision Summit. This talk introduces a new method for object detection for consumer goods and other applications based on measuring an object’s illumination invariant fluorescent chroma. Chroma is a measure of the colorfulness of an object relative to a similarly illuminated white object. Reflective chroma changes with the illumination, making it a poor choice for object detection in different environments. Fluorescent chroma, however, is nearly illumination invariant, making it an ideal input for object detection algorithms if the fluoresced light can be separated from the reflected light. Childers describes simple designs of illuminators, cameras, and filters to achieve that separation, and shows the classification accuracy of the system to be 95-100% under different lighting conditions using the nearest neighbors algorithm with neighborhood component analysis. BASF has designed hundreds of unique colors and coatings for consumer goods and other items that can be used in this application and the system has many advantages compared to CNN object detection systems, as it does not require extensive training sets or an unoccluded view of the object.

Views

Total views

150

On Slideshare

0

From embeds

0

Number of embeds

2

Actions

Downloads

1

Shares

0

Comments

0

Likes

0

×