Skip to content

reproduce success #20

Description

@LukeLIN-web

VideoMind 👍 usable(一遍过)

  • arxiv 2503.13444,ICLR 2026 accepted
  • 仓库:https://github.com/yeliudev/VideoMind ,last commit 6721d25 "Fix file name"(2026-02-08,3 个月前)
  • 测试时间:2026-05-19,GPU CUDA_VISIBLE_DEVICES=1(H100)

数据集

  • yeliudev/VideoMind-Dataset HF 公开,覆盖 27 个 grounding/QA benchmark(Charades-STA, NExT-GQA, QVHighlights, MLVU, …)
  • Annotation 文件 ~MB 级,免费下载
  • 视频两种格式:videos.tar.gz(原始,单数据集 ~50 GB+)/ videos_3fps_480_noaudio.tar.gz(压缩,单数据集 ~10 GB)。本次没拉全套
  • 用 HF Space yeliudev/VideoMind-2B 上的 demo 视频(~6 MB/个)验通路

环境踩坑

0 个。Conda env videomind (Python 3.11) + pip install -r requirements.txt 一遍过。

  • torch 2.4.0+cu121
  • transformers 4.45.2
  • peft 0.14.0
  • decord 0.6.0
  • deepspeed 0.15.4
  • nncore 0.4.5

推理验证

  • HF 拉模型:yeliudev/VideoMind-2B(1.9 GB,LoRA adapters for planner/grounder/verifier + 小 regression head)+ base Qwen/Qwen2-VL-2B-Instruct(4.2 GB)
  • 加载时间:9.3s(H100)
  • 自写 test_inference.py 跑 grounder role 单样本

输出片段

Video duration: 63.43s
Query: 'Why did the old man stand up?'
Generated in 1.5s

=== GROUNDER RESPONSE ===
The relevant moment happens in <|reg|>.
=========================
Grounded moments (top 5):
  [0] 00:00:20 - 00:00:28  conf=0.403
  [1] 00:00:04 - 00:00:13  conf=0.357
  [2] 00:00:16 - 00:00:25  conf=0.355
  [3] 00:00:23 - 00:00:30  conf=0.352
  [4] 00:00:08 - 00:00:16  conf=0.351

符合论文:文本响应嵌入 <|reg|> token,路由到 regression head 输出 ranked normalized [start, end] + 置信度(clamp 到视频时长)。Top-1 (20-28s) 与"老人站起来"语义吻合。

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions