LLLMWIKI
正文研究地图阅读入口元数据

来源笔记 · 更新于 2026-07-12

Broadcast2Pitch从非受控足球转播视频重建比赛状态

#near-cvpr-2025#video-understanding#tracking#benchmarking#primary-source

Broadcast2Pitch：几何与身份联合的足球比赛状态重建

会议：WACV 2026
全文深读：analysis.md

一句话结论

Broadcast2Pitch 用 joint line/circle/keypoint homography、LLaMA-3.2-Vision identity cues 与 ID-aware tracklet refinement，把 SoccerNet-GSR 的 GS-HOTA 从原 baseline 22.26 提升到 61.48；相对强 follow-up 的 55.82 增益为 5.66 点，代价是更复杂且未报告端到端 latency 的模块栈。

系统管线

  1. YOLOX detection；
  2. DeepEIoU + OSNet tracking/ReID；
  3. EfficientNet-attention U-Net 检测 pitch keypoints/lines；
  4. Lines/circles/keypoint fallback 联合优化 homography；
  5. LLaMA-3.2-Vision 从 crop 生成 role/jersey/color；
  6. IDATR split/merge tracklets；
  7. Majority voting/filtering；
  8. Jersey color clustering + mean pitch x 推断 team。

VLM 只处理 athlete crops 的身份属性，整场动作与战术不在任务内。

数据与训练

  • SoccerNet-GSR train/test；
  • YOLOX：SoccerNet-Tracking；
  • OSNet：SportsMOT；
  • Geometry：25 epochs，384×384，lr 1e-4；
  • LLaMA-3.2-Vision：SoccerNet jersey dataset 1 epoch，lr 1e-5；
  • 单 RTX 4090 24GB；
  • 未报告 end-to-end FPS、各模块延迟和总 GPU-hours。

核心结果

系统GS-HOTAGS-DetAGS-AssAIDF1
GSR Baseline22.2610.6746.46—
Golovkin et al.55.8241.6774.86—
Broadcast2Pitch61.4848.4778.0064.20

Field registration：

  • JaC5/JaC10：69.38/92.84；
  • MRE/MedRE：4.47/2.21 pixels；
  • Completeness：100%。

这些是 image-plane field-line reprojection 指标，不是 athlete pitch-coordinate meter errors。

消融

设置GS-HOTAIDF1
Keypoint homography48.2351.68
Line/circle optimization56.3961.31
Joint + keypoint fallback58.5161.57
+ IDATR61.4864.20

Identity：specialized PRTReID+OCR+ResNet 18.11 GS-HOTA；CLIP 60.13；LLaMA-3.2-Vision 61.48。LLaMA 相对 CLIP 增益较小，计算成本更高。

局限与风险

  • Team assignment 依赖 jersey color 与平均 x-position；
  • Tracklet filtering 可能漏掉困难对象；
  • VLM 可能生成无效/幻觉 identity tokens，缺格式失败率和 calibration；
  • 错误 identity cue 会导致 IDATR 错误 split/merge；
  • GS-HOTA 使用 5m tolerance；
  • Benchmark 内训练/测试，跨联赛与转播域未知；
  • Pipeline 没有 uncertainty propagation；
  • 缺端到端 latency、在线协议和人工复核流程。

对羽毛球 demo 的影响

可迁移 joint court geometry、identity-aware tracklet refinement、oracle ablation 与分层指标。羽毛球无需 jersey VLM/team heuristic，应改为 player click/re-ID、court side、shuttle/racket/contact。坐标 tolerance 应远小于 5m，并分 fixed/mobile camera 与报告全流程 latency。

证据评级

  • Soccer GSR：A-
  • Field registration：A-
  • VLM identity refinement：B+
  • 羽毛球迁移：C+

原始材料

  • raw/ingest/2026-05-23-broadcast2pitch-game-state-reconstruction/paper.pdf（11 页；有轻微 PDF syntax warnings）
  • raw/ingest/2026-05-23-broadcast2pitch-game-state-reconstruction/paper-text.md（1,259 行）
  • raw/ingest/2026-05-23-broadcast2pitch-game-state-reconstruction/analysis.md

相关页面

元数据
{
  "id": "2026-05-23-broadcast2pitch-game-state-reconstruction",
  "type": "source",
  "title": "Broadcast2Pitch：从非受控足球转播视频重建比赛状态",
  "status": "reviewed",
  "created": "2026-05-23",
  "updated": "2026-07-12",
  "venue": "WACV 2026",
  "published_at": "2026-03-23",
  "ingested_at": "2026-05-23",
  "tags": [
    "near-cvpr-2025",
    "video-understanding",
    "tracking",
    "benchmarking",
    "primary-source"
  ],
  "note_status": "reviewed",
  "source_type": "paper",
  "authors": [
    "Yin May Oo",
    "Yewon Hwang",
    "Muhammad Amrulloh Robbani",
    "Vanyi Chao",
    "Ankhzaya Jamsrandorj",
    "Hoang Quoc Nguyen",
    "Kyung-Ryoul Mun",
    "Jinwook Kim"
  ],
  "canonical_links": [
    "https://openaccess.thecvf.com/content/WACV2026/html/Oo_Broadcast2Pitch_Game_State_Reconstruction_from_Unconstrained_Soccer_Videos_WACV_2026_paper.html",
    "https://openaccess.thecvf.com/content/WACV2026/papers/Oo_Broadcast2Pitch_Game_State_Reconstruction_from_Unconstrained_Soccer_Videos_WACV_2026_paper.pdf",
    "https://github.com/yinmayoo185/SoccernetGSR"
  ],
  "raw_entry": "raw/ingest/2026-05-23-broadcast2pitch-game-state-reconstruction/",
  "analysis_note": "raw/ingest/2026-05-23-broadcast2pitch-game-state-reconstruction/analysis.md",
  "topics": [
    "topics/sports-ai-video-understanding",
    "topics/sports-ai-roadmap",
    "topics/video-understanding"
  ],
  "entities": [
    "entities/sportsmot"
  ],
  "claims": [],
  "questions": [
    "questions/question-badminton-stroke-correction-demo"
  ]
}