1Taking photos is freeFarmer wants to know disease type affecting tomato crop.Photos are from the same plant.Dataset: 80K plant leaf images of different crops https://plantvillage.org/No image pre-processing performed. No textual data used.Feature vectors generated automatically with deep learning.#1 #2 #3www.xyggy.com © 2019
View Slide
2Query with photo #1Top 5 results:1 potato_early_blight2 tomato_early_blight3 potato_early_blight4 potato_early_blight5 tomato_early_blight2nd result identifies diseasewww.xyggy.com © 2019
3Query with photo #2Top 5 results:1 tomato_early_blight2 celery_early_blight3 tomato_early_blight4 potato_early_blight5 tomato_early_blight1st result identifies diseasewww.xyggy.com © 2019
4Query with photo #3Top 5 results:1 soybean_leaf_blight2 celery_early_blight3 potato_early_blight4 tomato_early_blight5 soybean_leaf_blight4th result identifies diseasewww.xyggy.com © 2019
5Real world data is noisy3 queries with different photos from the sametomato plant give different predictions.Real world data is noisy.Let’s try with Thingy using multiple photos perquery.www.xyggy.com © 2019
6Query with photo #1 and #2Top 5 results:1 tomato_early_blight2 tomato_early_blight3 celery_early_blight4 potato_early_blight5 potato_early_blighttop 2 positions identify diseasewww.xyggy.com © 2019
7Add photo #3 to querytop 3 positions identify diseaseTop 5 results:1 tomato_early_blight2 tomato_early_blight3 tomato_early_blight4 potato_early_blight5 potato_early_blightwww.xyggy.com © 2019
8Taking photos is freeTaking photos is free. Ditto audio, video, sensorand other item data.Multiple items per query gives Thingyadditional information to work with to improverelevance.Perfect for a noisy world.www.xyggy.com © 2019
9Thingywww.xyggy.com © 2019