飞桨学习赛：乒乓球时序动作定位

Who this competition fits

适合已经会用 PaddlePaddle 2.0+ 训练端到端深度学习模型，并具备视频时序建模或动作定位经验的个人开发者。参赛者应能处理以 pkl 保存的连续帧 ppTSM 特征，而不是直接从原始视频训练。

What this competition gives you

There is a baseline you can run as-is — starting from it is usually far faster than building from scratch.

• BaselineForkable Paddle baseline project
• TutorialDownloadable baseline video
• BaselineOpen-source Paddle temporal action localization solution
• Data912 video feature files with training and test splits
• Reference paperBMN temporal action proposal paper
• Reference paperTSM efficient video understanding paper

From registration to a first submission

1. 01

   PaddlePaddle 2.0+

2. 02

   时序动作定位

3. 03

   视频特征建模

4. 04

   ppTSM特征处理

5. 05

   JSON/ZIP结果文件生成

6. 06

   AUC与时序提案评估

Before you commit: 难点在于从连续帧特征中定位极短、细微的挥拍片段：部分动作不到 1 秒，且小幅挥拍与非挥拍动作可能只在少数帧上有差异。还必须按规定生成 submission.json 并压缩为 submission.zip，且不能使用外部数据。