Supervision 0.26.0 is out 🔥🔥🔥. Our biggest update in months—packed with new features, support for VLM object detection, smarter and optimized annotators, and a fresh docs revamp!. 🚀 pip install -U supervision. 🔗 release notes → 🔗 docs →

RT @skalskip92: supervision, the open-source library I created 2 years ago, is crossing 30,000 stars on GitHub! . thank you to everyone wh….

RT @skalskip92: updated open-source object detection model leaderboard. 64 checkpoints; key metrics like mAP and F1, including small object….

🚀 RF-DETR also converges faster than other detectors when fine-tuning. check out the notebook to fine-tune RF-DETR by @skalskip92 . 🔗 📚 → . (3/3)🧵🏁

> the nano model scores 11 mAP higher than YOLO11-n on COCO mAP50:95, while being faster! . > both the small and medium variants of RF-DETR outperform even the biggest YOLO11 model on COCO mAP50:95, while being significantly smaller and faster! . 🔗 benchmarking notes →

RT @skalskip92: my project just crossed 30k stars on GH

RT @skalskip92: we released three new RF-DETR model sizes: nano, small, and medium. perfect of mobile devices. each model is the fastest an….

RT @AlexBodner_: Just released with @skalskip92: Detect the 3 second violation in NBA videos with AI. Featuring the @roboflow blog and ope….

supervision has been trending on @github 🔥. supervision comes with first-class support for pose and key-points annotation:. > effortlessly annotate keypoint skeletons.> assign custom labels to vertices.> support for ViTPose on @huggingface .> support for @ultralytics YOLO pose

RT @skalskip92: supervision has gained almost 1k stars since the release on wednesday; so cool. link: https://t.co/….

RT @roboflow: RF-DETR release next week. more SOTA models coming in smaller sizes, all open source and Apache 2.0.

@alibaba_cloud @moondreamai @GoogleDeepMind @huggingface @padillaRafa @SergioPaniego @onuralpszr 🙏 Grateful to all the contributors who made this release possible. @skalskip92 @onuralpszr @padillaRafa . (7/7)🧵🏁

@alibaba_cloud @moondreamai @GoogleDeepMind @huggingface @padillaRafa 📣 Special shoutout to @SergioPaniego and @onuralpszr for creating a @huggingface space for comparing detection capabilities of Qwen2.5-VL and Moondream with supervision-0.26.0 result parsing and visualization!. 🔗 🤗 → (6/n)🧵👇

@alibaba_cloud @moondreamai @GoogleDeepMind @huggingface the mean average precision metric has been re-implemented by @padillaRafa. mAP results now match pycocotools, and the interface is much simpler and easier to use for benchmarking any dataset or model. updated object leaderboard with mean average precision numbers calculated with

@alibaba_cloud @moondreamai @GoogleDeepMind Supervision now has support for pose estimation models from @huggingface transformers, so you can now visualize ViTPose and ViTPose++ skeletons directly in supervision using `sv.KeyPoints.from_transformer`. (4/n)🧵👇

@alibaba_cloud @moondreamai @GoogleDeepMind Supervision now supports visualizing and tracking objects with dense caption labels. > improved `sv.LabelAnnotator` with a “smart mode” where labels never overlap and always stay visible. > support for text wrapping and line breaks, so even long text prompts are displayed

VLMs are making real progress in detection and segmentation, and supervision-0.26.0 lets you easily parse and visualize their prediction results. > Support for parsing and visualizing detection results from @alibaba_cloud Qwen 2.5 VL, @moondreamai, and @GoogleDeepMind Gemini 2.0

RT @skalskip92: I'm back working on basketball AI. detecting players in paint; still need to smooth out the rough edges .

overall, SAMWISE proposes a lightweight extension of SAM2, endowing the base model with natural language-guided segmentation while tackling the key challenges of adapting SAM2 for text-guided online video segmentation and outperforming all previous methods with less than 5M

with <5M new parameters on top of SAM2 weights, SAMWISE delivers the top J&F on every benchmark while remaining markedly lighter than prior winners on MeViS, Ref-Youtube-VOS, and Ref-Davis benchmarks. (9/n) 🧵👇