How Bobtail works
Bobtail analyses a bitmap image and produces a structured description of everything in it. This page explains what that description is, what properties it has, and why those properties matter. It is written for an engineer who wants to know whether the approach is legitimate.
Decomposition, not recognition
Bobtail does not recognise objects, and it does not rely on training data or memorised patterns. It decomposes the image: it maps out the core features in the image, then the color transitions between them, and assembles this into a vector-based graph.
The result is the analysed structure of the image as vector geometry: the cores, the meridians that connect them, the contours around them, and the surfaces they bound. This works on photographs, not just logos and line art.
Never throw away information
Every stage of the decomposition is constrained to preserve the underlying data rather than simplify it away. Higher levels of the hierarchy add information, such as structural groupings, relationships and derived statistics, without removing access to the levels below.
A concrete consequence: the direct vectorization output is produced per color channel (red, green, blue and luminance), one vector definition each. These are deliberately not merged into one, because merging four channels into one loses information. If you need a single layer, you can collapse them yourself, knowing exactly what you are discarding.
Verifiable enhancement
Because the structure is preserved, it can be re-projected back into the pixel domain at any resolution. That gives upscaling almost for free: edges stay edges at any scale, and nothing is invented.
The claim we make about this is verifiability, not that pixels are unchanged. The process is deterministic, and the structure extracted from the output can be audited against the structure of the input. That is the guarantee a court or a radiologist needs, and one that no generative enhancer can make, because generative approaches are bound by a trade-off between perceptual quality and fidelity to the source.
Hierarchical representation
In development. A multi-resolution representation in which each coarser layer is a defined structural refinement of the layer below, so that the whole representation is invertible and verifiable. This is the substrate for structural editing: move an edge and the fill follows, select a region by what it is rather than by its pixels.
This part of the work contains open research, in particular on the pattern taxonomy and the layered data model, and we present it as such. See the roadmap for where it sits.
Pluggable interpretation
The output is not tied to a single recognition or editing method. Classical algorithms, learned models, or hybrid systems can consume the same decomposition. This lets the interpretation layer evolve independently of the representation, and lets each downstream task use whichever approach fits it best.
Whether the structure also makes a better input for neural networks than raw pixels is an open question that we intend to measure, not a claim we make today.
What Bobtail is not
It is not a replacement for neural networks in vision generally, and it does not try to be. It is a representation that makes some tasks tractable which are hard or impossible to do faithfully with generative methods, and that other methods can build on.