AIToday
Open-Source AIQiita 機械学習Published: Nov 21, 2026, 01:00 JST

auroray: 200,000-param MLP boundary filter, zero human labels

Send on LINE
auroray: 200,000-param MLP boundary filter, zero human labels

3 Key Points

  1. What happened

    An independent developer published auroray, a roughly 200,000-parameter MLP filter that draws candidate object-boundary lines on images. It is trained on zero human labels, using only programmatically generated geometry data, and is positioned as an AI-based alternative to OpenCV.

  2. Why it matters

    Because the filter only draws boundary candidates and knows no object categories, it can catch boundary candidates even on objects it has never seen, according to the developer.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

auroray sits deliberately against the current grain of very large general-purpose models. The developer's stated underlying question is how small a large network can be made, citing concerns around power consumption, compute cost, difficulty of transferring between models, and dependence on giant infrastructure and very large companies. A tiny, function-specific filter, the argument goes, is easier to embed in existing systems and easier to tune or retrain where large compute is not available.

The core V1Filter takes a 9×9×3 HSV patch and outputs a single boundary score for the center pixel. Training data is generated rather than labeled by hand: clean boundary patterns are created first and then degraded — lines broken or recolored, noise mixed in, patches partially erased — to build robustness. Negative cases explicitly teach the filter not to fire on boundaries that miss the center or do not connect edge to edge. The developer notes the filter can also pick up lines that are not object boundaries, such as wrinkles or patterns on clothing, and says strengthening boundary-candidate features was prioritized first.

A comparison tool pits V1Filter against OpenCV output, and a GUI tool overlays extracted boundary lines on a captured region in real time. Looking ahead, the developer wants to extend the same lightweight approach to overlap detection and centerline extraction, and eventually to replacing parts of CNN layers and building a general image classification system that pairs small filters with a vector database for easier additional training.

FAQ
What makes auroray different from OpenCV?
OpenCV relies on human-designed formulas for boundary extraction, while auroray lets the AI itself learn how to distinguish boundaries. The developer says this can surface distinguishing methods humans did not think of and makes it easier to add exception patterns as training data.
What kind of model is auroray?
Its core V1Filter uses a multi-layer perceptron, not a CNN, taking a 9×9×3 HSV patch as input and outputting a boundary score from 0.0 to 1.0 for the center pixel. The model has roughly 200,000 parameters.
How fast is auroray?
It supports stride=1 inference through GPU parallel processing, so it runs fast on high-resolution images on GPU. On CPU environments, processing time varies with image size and stride.
Qiita 機械学習Read Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleICLR peer-review overlap jumps from 12.0% to 21.3% as LLM use rises