← 回總覽

SentrySearch:基于 AI 的自然语言视频搜索

📅 2026-04-06 11:31 Nav Toor 人工智能 2 分鐘 1414 字 評分: 87
SentrySearch AI 计算机视觉 Gemini 开源
📌 一句话摘要 SentrySearch 是一款开源工具,利用 Gemini Embedding 模型实现对原始视频素材的自然语言搜索,无需进行视频转录。 📝 详细摘要 SentrySearch 解决了在海量原始视频片段中搜索特定内容的难题。它通过 Google 的 Gemini Embedding 模型,将视频和文本映射到同一个 768 维的向量空间中。这种方法允许用户直接对原始像素进行语义搜索,无需依赖元数据或转录文本。该工具开源,支持特斯拉哨兵模式(Sentry Mode)及通用 MP4 文件,并可通过 Qwen3-VL 支持本地模型,实现完全离线搜索。 📊 文章信息 AI 评分:

🚨 47 hours of footage. One sentence. The exact clip. Someone made it possible to Google Search your own videos.

It's called SentrySearch.

Type "red truck running a stop sign." It finds the exact 30-second clip from hours of raw footage. Instantly.

Not transcription. Not frame-by-frame scanning. Actual AI that understands what's happening in raw video.

No manual scrubbing through hours of footage. No timestamps. No guessing.

Here's how it works:

→ Index your video footage with one command

→ AI embeds raw video directly into a searchable vector space

→ Type any description: "white SUV cutting me off" or "pedestrian crossing at night"

→ It matches your text against actual video content. Not metadata. Not transcripts. Raw pixels.

→ Returns the top matches ranked by relevance

→ Auto-trims and saves the clip from the original file

Here's the wildest part:

There's no transcription. No frame captioning. No text middleman. Google's Gemini Embedding model projects raw video and text into the same 768-dimensional vector space. A sentence and a video clip become directly comparable. That's what makes sub-second search over hours of footage possible.

Every Tesla owner who's spent an hour scrubbing through Sentry Mode footage looking for one moment now has their tool.

Works with any MP4 footage. Not just Tesla. Also supports local models via Qwen3-VL for fully offline search.

Open Source.

查看原文 → 發佈: 2026-04-06 11:31:15 收錄: 2026-04-06 16:00:53

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。