Thx! Downloading all the photos took the most time. Object detection process took only several days.

At TripOffice, we use simple and widely-used tools: Python, NextJS, and MySQL.

Thx, see here: https://docs.mapbox.com/mapbox-gl-js/example/cluster/

I use MapBox, but similar things can also be done on Google Maps.

I thought it was ASICs?

CPU to GPU to FPGA to ASIC

All the acronyms

I haven't checked it, but such data should be easy to get from our database.

We manually labeled almost 1,000 chairs in various photos to train the model.

At first I wanted to use Google's Vertex AI for this purpose, but the costs were too high.

Besides, in this case, we would also have to upload 40 million photos to the cloud for Google to evaluate what’s on them.

YOLO is the best for such tasks; it works locally and is really fast.

This is not scraping. Many hotel sites share photos with their partners, we work with Lexyl (HotelPlanner).

All major booking sites share hotel photos with their partners.

We are in several partner programs, but now we mainly work with HotelPlanner.

I tried to detect models of specific chairs, but it's very difficult. To train the AI model, you need many photos of the same type of chair, and on the internet, you can usually find only one stock photo of each chair.

Make synthetic dataset with reposing

In our hotel photo database, we have everything: rooms, lobbies, bathrooms, pools, exterior shots of buildings, etc.

I trained an AI to recognize ergonomic chairs, but sometimes there were errors. For example, a chair in a hotel's SPA was always identified as an ergonomic chair. That's why we manually reviewed all 50k photos to verify them.

This desk is pretty useless, despite having an office chair: https://www.tripoffice.com/nigeria/lekki-ng/souz-suites-apar...

> I trained an AI, but sometimes there were errors.

I edited that down to a proper summary of AI in general

50k manually? How long did it take you?

I don't know, I assigned this task to trusted specialists from India.

Wow, manually reviewing 50k photos is a lot! Would you be willing to share what the cost of that was?

I created an app similar to Tinder that facilitates manual verification. Around 60 photos can be verified in one minute. The whole process took about a week and didn't cost much.

Sounds like YOLO verified the work of the specialists too! (;->

Doesn't sound that much. When I was playing with datasets, for simple tasks I only took around 3 seconds to classify an image. That's 1200/hour or on the order of 40 hours of work. That can't cost much when outsourced.

RLHF in the wild, nice.