Basic Patterns Reverse Search Engines
Objects, density maps, curves, surfaces, waveforms, feature hierarchies search by examples.
Gone are the days when text search was sufficient. Searching for similar strings of characters could be said to be "string fitting" problem, fundamentally the same problem as "curve fitting", and really, nearly any feature set fitting problem.
Unfortunately, Google Image search "Reverse Image Search" has been degraded from "deep visual feature matching", to "semantic label matching", making it much less useful in real world scenarios (to those who have no codenames for very specific features), because many visual problems that people care about do not map well into semantic feature space: there's no words for many colors, there's no words for many intricate patterns, and the best way to explain what pattern we meant, is by providing sample of that pattern.
Being a monopoly, unsurprisingly, Google has no need to provide advanced searches, like:
- 2D/3D/4D/ND voxelated/pixelated density maps search
- 2D/3D curves search
- 3D surfaces search
- Waveform matching search (surely, used to match and recognize all kind of signals)
- Machine-Learned feature hierarchies of object shapes for public to navigate and explore and reuse.
These searches are very obvious and likely had been developed by various groups for research needs, but I couldn't find versions of them for the general public.
In practice, the idea here is that these reverse-searches, while quite straightforward to implement, are areas ready for disruption, as they would be great enablers for human creativity based on pattern research. Just imagine, if people could create videos by looking up for curves they draw, or dance patterns, or phase spaces of hinge mechanisms, or dynamic deep feature overlays, transclusions, and so on.
To implement something like this, the most straightforward way would be to look, what file formats people already use to define these kind of objects, and create generalized formats that they all could be converted to, and then use those files as large training sets.