---
type: "post"
id: "rogii-wellbore-curves-recognition-0b7a29"
slug: "rogii-wellbore-curves-recognition-0b7a29"
title: "ROGII Wellbore curves recognition"
description: "Kaggle is hosting ROGII Wellbore curves recognition, a competition to turn scanned well-log images into separate traced curves."
retrieval_nugget: "Kaggle is hosting ROGII Wellbore curves recognition, a competition to turn scanned well-log images into separate traced curves. Its detailed Curve-PQ metric, baseline notebook and validation recipe make the challenge reproducible, while the $10,001 prize pool gives applied-ML teams a concrete reason to test methods for recovering legacy industrial data."
published_at: "2026-11-18"
updated_at: "2026-11-18"
record_date: "2026-11-18"
date_kind: "published_at"
topics: []
entities: ["Kaggle","ROGII Wellbore curves recognition","Pizzaboi","Igor Kuvaev","Dmitrii Rudenko"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "A new Kaggle competition asks participants to trace curves from scanned well-log images as continuous polylines."
prior_refs: []
source_url: "https://www.kaggle.com/competitions/rogii-2-wellbore-curve-tracing"
source_urls: ["https://www.kaggle.com/competitions/rogii-2-wellbore-curve-tracing"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:rogii-wellbore-curves-recognition-0b7a29"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/rogii-wellbore-curves-recognition-0b7a29/","markdown":"https://newruntime.com/posts/rogii-wellbore-curves-recognition-0b7a29.md","json":"https://newruntime.com/posts/rogii-wellbore-curves-recognition-0b7a29.json"}
markdown_url: "https://newruntime.com/posts/rogii-wellbore-curves-recognition-0b7a29.md"
---

# ROGII Wellbore curves recognition

## Retrieval answer

Kaggle is hosting ROGII Wellbore curves recognition, a competition to turn scanned well-log images into separate traced curves. Its detailed Curve-PQ metric, baseline notebook and validation recipe make the challenge reproducible, while the $10,001 prize pool gives applied-ML teams a concrete reason to test methods for recovering legacy industrial data.

## Summary

Kaggle is hosting ROGII Wellbore curves recognition, a competition to turn scanned well-log images into separate traced curves. Its detailed Curve-PQ metric, baseline notebook and validation recipe make the challenge reproducible, while the $10,001 prize pool gives applied-ML teams a concrete reason to test methods for recovering legacy industrial data.

## What happened

The Kaggle competition asks participants to process scanned well logs and return each annotated curve as a polyline: a sequence of source-image pixel coordinates. The competition description says historic logs often survive as scans of printed charts, while the values needed for comparison, field models and machine-learning datasets are not directly available as numbers. Operators currently trace lines manually. The task is therefore to recover curves from images containing multiple tracks, grids, labels, fills and other printed details.

The challenge is not simply to label curve pixels. Logs can contain thin, overlapping curves with different colors and solid, dashed or dotted styles. A line may cross another, leave a track and return, or be interrupted. The evaluation is also limited to zones where professional annotators traced curves: those annotators did not mark every real curve on every image. The metric penalizes invented curves inside a worked zone, but ignores predictions outside it. This makes the output structure and the distinction between separate curves important, not just pixel coverage.

Kaggle evaluates submissions with Curve-PQ, a panoptic-quality-style score computed per well and averaged across wells. Predicted and annotated polylines are resampled at one-pixel intervals and clipped to rectangular evaluation zones; pieces shorter than 50 pixels in-zone are dropped. A predicted and annotated curve are compared using precision and recall within a four-pixel tolerance, then matched one to one using the Hungarian algorithm. A pair counts only when its F-score is at least 0.5. The score accounts for matched, missed and spurious curves. Wells are divided into aligned windows of 1,600 rows; a well's score is the mean across windows containing at least one annotated curve, and the final score is the mean across wells. The metric implementation is included in the competition files.

The organizers point to an end-to-end baseline notebook for two Kaggle GPU T4s. It trains a U-Net on crops, then uses skeletonization and curve linking; it scores about 0.44 on a local validation set of 60 wells from batch 12. A full training run takes about six hours, within Kaggle's 12-hour session limit. The competition page also describes a local validation recipe using `train_curves_b12.csv` and `metric.py`, and says a complete test-set scoring run takes under a minute on a Kaggle machine.

Submissions contain one row per well, with curves encoded as JSON lists of `[x, y]` points; `null` marks a gap. Limits include 300 curves and five million points per well, with coordinates up to 1,000,000. Kaggle lists $3,001 for first place and $1,000 each for places two through eight, for a $10,001 total. Winners on the private leaderboard must provide code that reproduces their submission and a solution description. Data use is restricted to non-commercial competition participation, forum activity, academic research and education.

## What the archive adds

We have not covered this before; it opens a new thread.

## Trend signal

This does not yet fit a pattern New Runtime tracks. The competition concerns a specific image-to-curve reconstruction task, rather than a broader change in agent systems, interfaces or deployment practice.

## Why it matters

The task exposes a practical shape of industrial digitisation work: recovering structured measurements from documents that were designed for people to read. A model that detects the right pixels but merges crossing lines or fragments a continuous trace may still produce unusable data. For teams building extraction systems, the challenge makes the gap between image segmentation and semantically useful output explicit: downstream users need distinct, inspectable curves, not merely a mask.

The public metric and baseline also lower the cost of comparing approaches. Participants can reproduce a local score before submitting, inspect the matching rules, and see how tolerance, missed traces and false positives affect results. That does not demonstrate that the approach will digitise logs reliably in operational settings, but it offers a bounded testbed for methods that could later be evaluated against real domain requirements. The non-commercial data restriction is consequential: a strong competition result does not itself establish permission for commercial deployment.

## Limits and uncertainty

The competition page's baseline score is explicitly a local validation result, not evidence of final leaderboard performance or production accuracy. The public leaderboard uses 30% of test wells and the private leaderboard uses 70%, so the public score is not the full evaluation. The page does not establish how well a high-scoring solution generalizes to other logging formats, image quality, annotation conventions or operational use. Nor does a prize-winning submission by itself demonstrate that manual digitisation can be replaced safely. Data use is limited to non-commercial purposes, and winners must provide reproducing code to receive prizes.

## Reader takeaway

This is a sharply defined applied-ML challenge, not a general-purpose AI release. Its most reusable detail is the evaluation design: separate curves are matched locally within 1,600-row windows, with explicit tolerance and penalties for missed or spurious traces. Treat the baseline's 0.44 as a starting point for experimentation, not as proof that the digitisation problem is solved.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
