LogoThread Easy
  • 探索
  • 線程創作
LogoThread Easy

Twitter 線程的一站式夥伴

© 2025 Thread Easy All Rights Reserved.

探索

Newest first — browse tweet threads

Keep on to blur preview images; turn off to show them clearly

一筐大模型揉一起做篮球识别!

给大家看这个识别效果,投篮位置,是否进球,球衣编号,球在哪里(甚至裁判手里拿的另外一个球都能捕捉到),篮筐,运动员都能识别。

总计用了这些模型:

F-DETR (检测球员)—— 这是个类 DETR 的实时目标检测器。微调后用来检测球员、球衣号码、裁判、篮球,甚至投篮类型。

 SAM2 (跟踪球员)—— 用于分割与跟踪。它在球员被遮挡后重新识别,并在身体接触中保持目标 ID 稳定。

SigLIP + UMAP + K-means (无监督球队聚类)— 结合视觉-语言嵌入与无监督聚类,通过统一的颜色和纹理将球员自动分组,无需人工标注

SmolVLM2 (识别球员号码)— 这个比较猛,今年2月份发布的,有256M, 500M, 2.2B 三个版本。一般用在OCR场景,是个VLM,经过 NBA 球衣裁剪图像微调后,识别队服和编号准确率从 56%提升至 86%

ResNet-32 — (号码分类)一种经典 CNN,经过微调用于球衣号码分类,测试准确率达到 93%,优于微调后的 SmolVLM2

原文写的非常棒,作为学习资料足够了,推荐给大家:

一筐大模型揉一起做篮球识别! 给大家看这个识别效果,投篮位置,是否进球,球衣编号,球在哪里(甚至裁判手里拿的另外一个球都能捕捉到),篮筐,运动员都能识别。 总计用了这些模型: F-DETR (检测球员)—— 这是个类 DETR 的实时目标检测器。微调后用来检测球员、球衣号码、裁判、篮球,甚至投篮类型。 SAM2 (跟踪球员)—— 用于分割与跟踪。它在球员被遮挡后重新识别,并在身体接触中保持目标 ID 稳定。 SigLIP + UMAP + K-means (无监督球队聚类)— 结合视觉-语言嵌入与无监督聚类,通过统一的颜色和纹理将球员自动分组,无需人工标注 SmolVLM2 (识别球员号码)— 这个比较猛,今年2月份发布的,有256M, 500M, 2.2B 三个版本。一般用在OCR场景,是个VLM,经过 NBA 球衣裁剪图像微调后,识别队服和编号准确率从 56%提升至 86% ResNet-32 — (号码分类)一种经典 CNN,经过微调用于球衣号码分类,测试准确率达到 93%,优于微调后的 SmolVLM2 原文写的非常棒,作为学习资料足够了,推荐给大家:

A coder, road bike rider, server fortune teller, electronic waste collector, co-founder of KCORES, ex-director at IllaSoft, KingsoftOffice, Juejin.

avatar for karminski-牙医
karminski-牙医
Mon Nov 03 22:37:43
now I really want a good ai first email client with strong CRM integration (that’s not a sales tool)

now I really want a good ai first email client with strong CRM integration (that’s not a sales tool)

VC by day @untappedvc, builder by night: @babyagi_, @pippinlovesyou @pixelbeastsnft. Build-in-public log: https://t.co/UdHHGbZba5

avatar for Yohei
Yohei
Mon Nov 03 22:37:31
6 dads vs. 1 hotel provided stroller

6 dads vs. 1 hotel provided stroller

Marketer, self-taught developer, and founder of @Bento and https://t.co/lcsIohchEv. Designing a quiet family life in 福岡, Japan. DMs open if you need email help 🌿

avatar for ˗ˏˋ Jesse Hanley ˎˊ˗
˗ˏˋ Jesse Hanley ˎˊ˗
Mon Nov 03 22:32:03
RT @jonsommet: @shl You have to choose safe or affordable.

RT @jonsommet: @shl You have to choose safe or affordable.

Founder/CEO @Gumroad

avatar for Sahil Lavingia
Sahil Lavingia
Mon Nov 03 22:30:25
RT @togethercompute: 📄New Guide: Running nanochat on instant clusters!  

Train and inference @karpathy's end-to-end ChatGPT clone on Toget…

RT @togethercompute: 📄New Guide: Running nanochat on instant clusters! Train and inference @karpathy's end-to-end ChatGPT clone on Toget…

Building @EurekaLabsAI. Previously Director of AI @ Tesla, founding team @ OpenAI, CS231n/PhD @ Stanford. I like to train large deep neural nets.

avatar for Andrej Karpathy
Andrej Karpathy
Mon Nov 03 22:30:10
RT @arvidkahl: Either this is a really great security process or a very devious way of getting people to sign up for an affiliate program.…

RT @arvidkahl: Either this is a really great security process or a very devious way of getting people to sign up for an affiliate program.…

Building https://t.co/od97B0HVrk and https://t.co/666FnyVVE0 in Public. Raising all the boats with kindness. 🎙️ https://t.co/6w69DZmi8H · ✍️ https://t.co/lpnor5rsTW

avatar for Arvid Kahl
Arvid Kahl
Mon Nov 03 22:29:23
  • Previous
  • 1
  • More pages
  • 1128
  • 1129
  • 1130
  • More pages
  • 2118
  • Next