Experimented with multi-gpu training of indoor object detector using RetinaNet and Open-Images – V5 dataset
The detector consists of 24 classes such as table, bed, sofas, home and kitchen appliances, etc.
The training ran on AWS P3.x large instances4
Nvidia V100 GPUS
244 GB CPU RAM
Training time – 5 hours for 10 epochs
The code will work on lower end GPUs as well as single GPU too. It’s a simplified wrapper on original retinanet.
Possible use-case for real-estate and housing rental businesses like Airbnb, Nestaway, etc
Usually it is the trend that while looking for houses to buy or rent customers find confidence in the ones having more pictures and amenities mentioned, which at present has to be done manually most of the times.
Such automated tagging could decrease the load on the seller. Auto tags based on
📌 how furnished the house is
📌 number o windows, doors, etc
📌 amenities such as swimming pools, kitchen and home appliances, parking areas.
More tagged the data, better is the search for customers looking for houses.