blobboards-descriptor

prittjam/blobboards-descriptor

blobboards-descriptor is a image feature extraction model from prittjam. Use it for the image feature extraction task on the model card, and read the license before you ship it in a product. It is set up for onnx. The card lists the license as mit.

Log-polar patch descriptor used by BlobBoards.jl to match detected blobs against a board gallery (findboards / trackboards).

Downloads · 30 days

Access

Public

Updated Aug 18, 2026

Repo size

8.3 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.onnx4.3 MB · 52%

At a glance

Task: Image Feature Extraction
Library: onnx
License: mit
Access: Public
Created: Aug 18, 2026
Updated: Aug 18, 2026
SHA: 3a7e2821

Model card

From the Hugging Face model README

BlobBoards descriptor network

Log-polar patch descriptor used by BlobBoards.jl to match detected blobs against a board gallery (find_boards / track_boards).

Files

file	sha256
`track_logpolar_circ_s96_iter4.onnx`	`6efc291ab94e812ce043d919ecc72ded7e4b16bd1c0c50b1492b1ce245bae8b9`
`track_logpolar_circ_s96_iter4.tar.gz`	`f92e3566ea747941f7efbd96ca0d5b322fdb74d2fcf5fdf7c8d582282aeaac67`

The .tar.gz wraps the same .onnx for Julia's Pkg artifact mechanism (git-tree-sha1 = 4018efc39186b1f7c287771d69da99817d73afd5); BlobBoards' Artifacts.toml pins it by commit URL + sha256.

Revisions

Consumers must pin a revision tag (or commit sha) — never main. BlobBoards' TOML loader refuses floating revisions by design.

tag	network
`iter4`	RealFourthIteration log-polar s96 export (2026-08-10). Zero `Pad` nodes, so ONNXRuntime's CUDA EP places the whole graph on the GPU (~65 µs/patch at batch 64).

Newer weights are published as new files + a new tag; existing tags are immutable, so a pinned config keeps resolving bit-identical weights.

Usage

In a BlobBoards run config (configs/*.toml):

[descriptor]
model = { artifact = "track_logpolar_circ_s96_iter4" }

descriptor_network_path resolves the name through the package's Artifacts.toml catalog: the tarball is downloaded once into Julia's content-addressed artifact store and verified against its pinned sha256.

Details

Input: 64×64 log-polar-canonicalized grayscale patches, batched; the "s96" in the name is the 96 px outer sampling radius of the canonicalization, not the input resolution.
Output: L2-normalized descriptor per patch; each patch is normalized independently and all BatchNorms are folded, so batch size is purely a throughput knob (numerically exact at any batch).
Export carries no Pad operators — no depad rewrite needed for the ONNXRuntime CUDA execution provider.

Task: Image Feature Extraction
Library: onnx
License: mit
Created: Aug 18, 2026
Updated: Aug 18, 2026