This project entails allowing the google cloud vision API to scan and return their key descriptions and probability scores. This will be step one of a multi-step process. If step one is a success, step two comes.
The client needs an automated program that loads photos of homes (varying quality and angles and stuff in the way) up to GCV and captures the key terms and probability scores for each.
The project will start only with exterior photos of the homes and the developer has to try to define architecture style and specific predefined features (this entails customizing what GCV looks for).
The photos that have to be used will not be labeled in most cases and perhaps only 1-3 will be exterior photos and the best interior. The client only needs the exterior photos for this first project. Upon success, they will want to start the next.
Features to be Integrated
The client has visually outlined some very key features of the architectural styles they wish to identify. GCV output is rudimentary in detail but does a pretty good job. It's their goal to hone its abilities for their customized needs and keep the customization non-public.
So they are just identifying exterior architectural styles and particular features of each using varying degrees of quality residential real estate photos. The photos will be pulled automatically from their database to GCV, the key terms and probabilities captured and returned for analysis via a proprietary in-house system. They want to customize what GCV looks for in the photos.
We have worked on different projects which involve the use of Google Cloud Vision. In one of the projects, we have used the face detection feature of Google cloud vision API. In that project, we were supposed to do the face recognition of a person entering the office premises being recorded by office cameras.
The major challenge faced for this project was identifying and matching the new results with the ones already saved as standards against which we had to check. For the same, we used the GCV APIs results and did workarounds based on its approximates.