用AI构建篮球投篮评估器：本地运行的计算机视觉投篮分析系统
摘要 该项目构建了一个本地篮球投篮追踪器，记录投篮命中、失手、投篮速度、投篮弧度、身体姿势和姿势。我们使用Roboflow训练的RF-DESTR物体检测模型定位球和篮筐，配合RF-DESTR关键点估计模型追踪射手姿势。整个系统运行在笔记本电脑上。 我从初中开始打篮球。然而，仍有一些下午我在公园里打针，感觉自己没有任何进步。我想用计算机视觉来解决这个问题。目标是创建一个分析工具，自动记录击球、失误、... 
评论
点赞
收藏Top Models for AI Data Labeling
SUMMary Model rankings shift with every release, so static lists go stale fast. This post explains how AI data labeling works, shows how to compare 25 current models on the Roboflow Playground object ... 
How to Build an AI Video Car Damage Inspector
Summary This tutorial builds a rental-car walkaround inspector. You upload a video of the car, and the system timestamps it before it even finishes uploading, runs an RF-DETR damage model inside a Rob... 
How Cloud Connected AI Products Run On-Prem
Summary The Purdue Model splits an industrial network into levels (physical process, PLCs, SCADA, MES, the DMZ, then enterprise IT) so a breach in business software never reaches the machinery, and it... 
Qwen3.8-Max for Vision: Benchmarks, Strengths, and Real-World Tests
This week, Alibaba released Qwen3.8-Max, its most capable Qwen model so far. The model has 2.4 trillion total parameters and uses a Mixture-of-Experts architecture, activating around 95 billion parame... 
How to Detect Small Objects in Drone Imagery
SUMMARY Small objects in drone imagery are detectable with a custom-trained model: this tutorial trains RF-DETR on roughly 7,000 aerial images with 120,000+ person annotations to find nine classes of ... 
Dwell Time and Zone Analytics with Vision AI
Summary Dwell time and zone analytics measure how long each person stays inside a defined area, a stronger signal than foot traffic alone. This tutorial builds the full pipeline in Roboflow Workflows:... 
P&ID Symbol Detection for Engineering Drawing Digitization
Summary You can automate P&ID symbol detection: an RF-DETR model trained on 3,800 annotated diagrams reaches 99.2% mAP@50 across 11 symbol classes, from gate valves to instrument tags. This tutorial b... 
Zero-Shot vs. Fine-Tuned Models: When to Train Your Own
SUMMARY Zero-shot models enable rapid prototyping without labeled data or training, while fine-tuned models provide higher accuracy, faster inference, and production reliability through additional lab... 
How Often Should You Retrain a Computer Vision Model?
Summary Retrain a computer vision model when production evidence shows it no longer represents the task: a measured performance decline, changed production images, repeated failure patterns, new requi... 
Human-Object Interaction Detection with RF-DETR
SUMMARY You can build human-object interaction detection without training an interaction model: train RF-DETR to detect workers, forklifts, pallets, and carts (77.5% mAP@50 here), then pass the annota... 
Computer Vision PLC Integration: Turn Detections into Machine Actions
Summary To connect computer vision to a PLC, run your model in a Roboflow Workflow that turns detections into a decision (reject_signal = true), then write that value to a PLC tag over OPC UA, Modbus ... 
Powder-Coat Defect Detection
Summary You can automate powder coat defect detection by training RF-DETR on a labeled coating dataset to find craters, orange peel, bubbles, and scratches (74.0% mAP@50 here), then having Gemini 2.5 ...
使用树莓派和Roboflow构建泳池监控器
摘要 如何利用树莓派、手机摄像头以及Roboflow分类模型，轻松判断泳池是否需要补水。跟着教程一步步操作，你只需一个周末就能亲手打造一台这样的设备。 我的同事迈克尔不断发布在边缘端运行的Roboflow演示视频，每次看到他的分享，我都会暗自感叹：我也想造出这样一款设备！于是，当我回到圣路易斯时，便把树莓派带在身边，从他的经验中汲取灵感。 是父亲提出了这个实际应用场景：我们的泳池会随着时间推移逐渐... 
Machine-Vision AI for Surface Defect Severity Assessment
Summary Defect severity classification means scoring each detection instead of just flagging it: pair a trained RF-DETR model with a Custom Python block that checks bounding-box area, so large defects... 
Shop Floor Data Collection: How to Automate It with Computer Vision
Summary Shop floor data collection is the process of recording what happens during production: unit counts, cycle times, defects, changeovers, downtime, and safety events. You can automate it with cam... 
What Is a Unified Namespace? Where Vision AI Fits in the UNS
SUMMARY A unified namespace (UNS) is a single hierarchical data model, usually carried over MQTT, where every system in a plant publishes its current state once and every consumer subscribes to the to... 
基于相机人工智能的生产线监控
Roboflow 的 AI1 Camera 结合 RF-DETR 模型和 Workflows，可将普通产线摄像头实时转为缺陷检测系统，实现瓶盖/标签缺失自动识别与计数，mAP@50 达 99.9%。文章详解从数据采集、模型训练到边缘推理落地的完整流程，提供可复现的工业视觉方案。 
如何利用计算机视觉为儿童足球比赛制作自动高光集锦
Roboflow 博客分享了一套利用计算机视觉自动剪辑儿童足球比赛精彩片段的完整方案。该方案在 Roboflow Workflow 中使用两个 RF-DETR 模型（球员/球/球门检测器和球衣号码检测器）配合 BoT-SORT 进行跟踪，并通过本地 Python 管道将检测结果转化为球衣号码、唯一追踪的儿童 ID、进球判定及最终的高光剪辑。此外，还提供了一个 FastAPI Web 应用，允许家长上传比赛视频，点击指定球员后自动生成包含该球员参与动作的剪辑合集。 
在 NVIDIA DeepStream 上运行 RF-DETR
Roboflow 提供在 NVIDIA Jetson Orin NX 上部署 RF-DETR 目标检测模型的详细教程。内容涵盖利用 DeepStream 构建 TensorRT 推理引擎、编写自定义解析器处理模型输出、以及配置 RTSP 流实时检测的具体步骤。文章附带基于 JetPack 7.2 的性能数据，适合需要进行边缘端视觉算法落地优化的开发者参考。 
利用计算机视觉检测汽车座椅 Hog Ring
Roboflow 发布基于工作流的汽车座椅 Hog Ring 自动化检测方案。利用 RF-DETR 进行目标计数，结合自定义 Python 逻辑与 Gemini 视觉审查，实现缺陷识别与 MES 集成。提供现成的标注数据集，适合工业视觉落地参考。 
Gemini 3.6 Flash 视觉能力评测：视频最强，但物体检测出现退化
Roboflow 对 Google 发布的 Gemini 3.6 Flash 进行了视觉能力实测。该模型在视频处理上表现最佳，且成本更低、速度更快，但在物体检测任务中出现退化，倾向于输出模糊的大框而非精确边界，甚至导致 JSON 解析失败。相比之下，其在计数和数据提取方面仍保持领先。对于依赖高精度目标检测的工程应用，建议谨慎升级或继续使用 3.5 Flash。 