Skip to content
View 8rulerstar's full-sized avatar

Block or report 8rulerstar

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please donโ€™t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this userโ€™s behavior. Learn more about reporting abuse.

Report abuse
8rulerstar/README.md

Hi there, I'm Byungkun Jang ๐Ÿ‘‹

Thanks for dropping by! I'm based in Korea and build computer vision and machine learning tools, and I'm especially interested in how much you can trust the measurements taken from pose estimation models. I also like making games.

ํ•œ๊ตญ์—์„œ ์ปดํ“จํ„ฐ ๋น„์ „๊ณผ ๋จธ์‹ ๋Ÿฌ๋‹ ๋„๊ตฌ๋ฅผ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ํŠนํžˆ ํฌ์ฆˆ ์ถ”์ • ๋ชจ๋ธ์˜ ์ธก์ •๊ฐ’์„ ์–ผ๋งˆ๋‚˜ ๋ฏฟ์„ ์ˆ˜ ์žˆ๋Š”์ง€์— ๊ด€์‹ฌ์ด ๋งŽ๊ณ , ๊ฒŒ์ž„ ๋งŒ๋“ค๊ธฐ๋„ ์ข‹์•„ํ•ฉ๋‹ˆ๋‹ค.

Highlights: pose estimation evaluation

Ultralytics (YOLO) ยท ultralytics/ultralytics#26564
Found that convert_coco(use_keypoints=True) kept people with no labelled keypoints, so YOLO pose validation counted them as targets. On a 200-image COCO val2017 sample, pose mAP50-95 for yolo26n-pose goes from 0.347 to 0.576, matching pycocotools. I reported and reproduced it in #26563, and the maintainers merged it as a smaller fix at the source.

supervision (Roboflow) ยท roboflow/supervision#2687
Added sv.metrics.KeypointMeanAveragePrecision, the first pose metric in supervision: COCO keypoint mAP using Object Keypoint Similarity for sv.KeyPoints. Its AP, AP50, AP75, APm and APl agree with pycocotools to within about 1e-8 on synthetic tests. Before that I fixed from_yolo reading pose labels as polygons (#2655, released in 0.30.8).

Projects

poseaudit ยท Python
Measures how far angles, tilts and lengths derived from pose keypoints disagree with the labels, not just whether the keypoints land near them. With YOLO11n-pose on 322 elbows from 200 COCO val2017 images, 58% of angles disagree with COCO's labels by 15ยฐ or more when the arm segments average under 30 px (37% at 30 to 60 px, 26% above 60 px). Part of that is the labels' own noise.
PyPI โ†’

Epokio ยท Python
Sends a phone alert when an ML training run on your machine or a GPU server stalls, hits NaN or finishes. It reads the logs your runs already write (Ultralytics, Hugging Face, Lightning, Keras, TensorBoard), with no code changes and no account. Web page, terminal view and a macOS menu bar app.
PyPI โ†’

(Also a small browser game: Stella Ball, Play on itch.io.)

More open source

  • sktime/sktime#11233: Re-enabled tests for two estimators that were skipped by mistake, a follow-up the maintainer requested while reviewing my other (still open) PR.
  • MODSetter/SurfSense#2027: Deleting the last connection to a host now also revokes its network access. Hosts still used by other connections keep it.

See all merged pull requests โ†’

If you're working on pose estimation, building games, or just want to talk about either, I'd love to hear from you. Questions and feedback are always welcome.

Contact: bkbkjang@gmail.com

Pinned Loading

  1. poseaudit poseaudit Public

    Audit how far off the angles, tilts and lengths a pose model measures are

    Python

  2. epokio epokio Public

    Phone alerts when an ML training run stalls, hits NaN or finishes, from the logs it already writes. No code changes, no account.

    Python 2