I wanted to share here what would be a minimal viable product for my project. In my case, it is not exactly a product since I am working on a research article. I am working on algorithms that produce 3D models of real-world scenes using images (for instance a bedroom scanned with a Kinect or a simple digital camera…), while identifying the categories of objects that populate the scene (this is a bed, this is a sofa, etc…). It is not so much about developing a complete pipeline that produce the best models ever, than actually trying new techniques and seeing how promising they are, in which case we write an article about it.
My current project now consists in proposing a method that is able to disambiguate and count 3D objects of similar category in a scene. Though the goal doesn’t seem too complicated, it is actually not so easy to solve when objects are very close to each other, or if no clear separation in general appears in the images (for instance when you have a series of cars parked along a street).
In order to have a minimum viable product, that is an article that could be proposed to conférences or journals, we need:
- Literature review on existing methods
- Clear description of the method we plan to test
- Code running on very simple data (proof of concept)
- Evaluation on real data