Project SmartKitchen/ Experiment 1: Image Recognition Prototype
From MozillaWiki
<< Connected Devices/Projects/Smart Kitchen
Contents
Status: DONE
Leap Of Faith Assumption (LOFA) We Are Testing:
(I) We assume that we can identify some percentage of food via image recognition
Hypothesis:
We believe that if we give a user a device for identifying their inventory that we will be able to recognize 20% of items with our existing training to date.
Experiment Steps:
- Give 3 internal Mozilla users (not on the team) our prototype device
- Have them sign up for a Kik account
- Instruct them on how to provision their wifi
- Have them try to identify items in their kitchen with the device when they are done with their shopping
- Collect metrics on the accuracy of the item identification
Experiment Timeline:
7/18 | Experiment Details Created |
07/19 | Email Sent out to SF/MTV Staff for 3 testers |
07/22 | Trial Instructions and Wifi Set Up Instructions completed |
07/25 | Software completed for prototype (server, username, etc.) |
07/27 | Hardware completed from Kevin and delivered to SF Office |
07/27 | Study Begins (each participants have 1 week and are starting on different days) |
07/27 | Metric Collection Begin |
7/28-8/5 | First week of metrics collected with no training |
08/06 | Upload additional training and update classifiers.json |
8/6-8/12 | Second week of metrics collected with additional training on common ingredients of trial participants |
Experiment Details:
Key Findings:
- Recognized 25% of the images but were very loose recognition. (i.e. Food)
- Need to look at another recognition survey
- Need to incentivize survey participants
- Team needs to look at actual images to see what the problem is.
Next Steps:
- Possible secondary experiment: comparing two IR services