Insights About Places with Deep Learning Computer Vision
summary
Uses 217,000 images with ~1.5M scenic votes to train CNNs that predict place beauty, showing natural scenes and characterful built elements score high while flat grass scores low. Applies transfer learning with MIT Places to scale from one image/km to dense London maps and to property work: 300M listings (avg 14 images each) get CV + text + metadata models that predict condition against surveyor labels, use active learning and ChatGPT-assisted labeling, and run in production for valuations and Beautiful Places CIC data products.
AI-written summary · about
chapters (8)
- 0:00Intro
- 1:25Using computer vision (CV) to predict the beauty of places
- 5:24Measuring beauty with AI
- 14:02Insights about properties using CV
- 23:52Using CV to understand consumer preferences
- 29:57Getting started with CV
- 32:56Beautiful Places
- 38:35Outro
related talks
11mTraining machines to see scenic beauty
20mPragmatic Deep Learning for image labelling
04mUsing Computer Vision to Understand Big Visual Data
01h 18mComputer Vision after the Victory of Data
19mHow we leverage AI to build great maps at massive scale
33mDeep Learning at Booking.com
46mML Use Cases at Think Big Analytics
02mWe painted the whole town