ML Kit の概要と Base API / first step of ML Kit

ML for Mobile Developers
https://gdg-tokyo.connpass.com/event/95203/

Transcript

1. ML Kit ͷ֓ཁͱ Base API Yuki Anzai @yanzm Google Developers Expert for Android ML for Mobile Developers
2. ML Kit ͱ͸ • Firebase ͷػೳͷҰͭ • ػցֶशΛར༻͢ΔػೳΛΞϓϦʹ؆୯ʹ૊ΈࠐΉͨΊͷϞόΠϧ SDK • ݱࡏ͸ β • iOS ͱ Android Ͱ࢖͑Δ • https:// fi rebase.google.com/docs/ml-kit/
3. on-device or in the cloud On-device Cloud Text recognition : ςΩετೝࣝ O O Face detection : إݕग़ O - Barcode scanning : όʔίʔυεΩϟϯ O - Image labeling : ը૾ͷϥϕϧ෇͚ O O Landmark recognition : ϥϯυϚʔΫೝࣝ - O Custom model inference : ΧελϜϞσϧਪ࿦ O -
4. on-device vs Cloud • on-device API • ϩʔΧϧͰಈ࡞ɺ଎͍ • Firebase ͕ػցֶशͷϞσϧΛ͋Β͔͡Ίμ΢ϯϩʔυͯ͘͠ΕΔ • Cloud API • αʔόʔͰॲཧɺߴػೳ • ωοτϫʔΫ઀ଓ͕ඞཁ
5. Pricing https:// fi rebase.google.com/pricing/ your account's fi rst 1000 Cloud Vision API calls/month are free
6. • Features • ը૾͔ΒςΩετΛೝࣝ • on-device API • ແྉ • શͯͷϥςϯจࣈΛೝࣝ • Cloud API • ຖ݄࠷ॳͷ 1000 API call ͸ແྉʢ1000+ ͔Β͸ैྔ՝ۚʣ • 50 Λ͑͜ΔݴޠΛೝࣝʢ೔ຊޠؚΉʣ Text recognition (OCR) : ςΩετೝࣝ
9. • Features • on-device API ͷΈ • إͷྖҬɺϥϯυϚʔΫʢ໨ɾ๹ɾඓɾࣖɾޱʣͷҐஔೝࣝ • إͷද৘ʢ໨ͷ։ด౓߹͍ɺসإͷ౓߹͍ʣͷೝࣝ • ಈըͷϑϨʔϜؒͰಉ͡إΛτϥοΫՄೳ • 100Ҏ্ͷ఺͔Βߏ੒͞ΕΔ2࣍ݩͷྠֲ৘ใʢإͷ֎ֲɾ໨ɾඑɾඓɾޱʣ Face detection : إݕग़
10. Face contour https:// fi rebase.google.com/docs/ml-kit/detect-faces
11. • Features • on-device API ͷΈ • ΄ͱΜͲͷඪ४ϑΥʔϚοτΛαϙʔτ • 1࣍ݩϑΥʔϚοτ : Codabar, Code 39, Code 93, Code 128, EAN-8, EAN-13, ITF, UPC-A, UPC-E • 2࣍ݩϑΥʔϚοτ : Aztec, Data Matrix, PDF417, QR Code • ࣗಈϑΥʔϚοτݕग़ • structured data ͷऔΓग़͠ • όʔίʔυͷ޲͖ʹΑΒͣݕग़Մೳ Barcode scanning : όʔίʔυεΩϟϯ
12. format : 256 valueType : 9 rawValue : WIFI:S:SB1Guest;P:12345;T:WEP;; displayValue : SB1Guest 12345 boundingBox : Rect(300, 457 - 669, 824) encryptionType : 3 ssid : SB1Guest password : 12345
13. • Features • ը૾ͷ಺༰Λղੳ͠ɺೝࣝͨ͠΋ͷͷϥϕϧΛ͚ͭΔ : ਓɺ෺ɺ৔ॴɺ׆ಈͳͲ • on-device API • ແྉ • 400+ labels Λαϙʔτ • Cloud API • ຖ݄࠷ॳͷ 1000 API call ͸ແྉʢ1000+ ͔Β͸ैྔ՝ۚʣ • 10,000+ labels Λαϙʔτ Image labeling : ը૾ͷϥϕϧ෇͚
14. label : Building con fi dence : 0.77894384 entityId : /m/0cgh4 label : Palace con fi dence : 0.75397676 entityId : /m/05zp8 label : landmark con fi dence : 0.9432406 entityId : /m/05_5t0l label : town con fi dence : 0.9333225 entityId : /m/0dx1j
15. label : Food con fi dence : 0.9649049 entityId : /m/02wbm label : Cuisine con fi dence : 0.91778296 entityId : /m/01ykh label : food con fi dence : 0.9399401 entityId : /m/02wbm label : cuisine con fi dence : 0.9263104 entityId : /m/01ykh
16. val options = FirebaseVisionLabelDetectorOptions .Builder() .setConfidenceThreshold(0.9f) .build()
17. • Features • ը૾͔Β༗໊ͳϥϯυϚʔΫΛೝࣝ • ϥϯυϚʔΫ໊ • ஍ཧ࠲ඪ • Knowledge Graph entity ID • ը૾಺ͰͷϥϯυϚʔΫͷྖҬ • ຖ݄࠷ॳͷ 1000 API call ͸ແྉʢ1000+ ͔Β͸ैྔ՝ۚʣ Landmark detection : ϥϯυϚʔΫೝࣝ
18. landmark : Amsterdam Centraal Railway Station con fi dence : 0.86155003 entityId : /m/0bbw52 locations : 52.378068, 4.899774 boundingBox : Rect(33, 504 - 956, 928) landmark : Amsterdam con fi dence : 0.5167069 entityId : /m/0k3p locations : 52.373811, 4.890951 boundingBox : Rect(187, 644 - 757, 843)
19. ΧελϜϞσϧਪ࿦ • Firebase Ͱ TensorFlow Lite ͷϞσϧΛϗεςΟϯά • Firebase SDK ͕Ϟσϧͷμ΢ϯϩʔυΛϋϯυϦϯά • Ϟσϧͷߋ৽΋Մೳ • apk ʹόϯυϧͨ͠ϞσϧΛ Firebase SDK ܦ༝Ͱར༻͢Δ͜ͱ΋Մೳ • on-device API ͱͯ͠ར༻