Video Table Extractor

Turn the tables shown on screen in your own video file (screen recordings, slide talks, product demos) into CSV and Excel files. Every row says the frame time it was read from and links to an image of that frame, so you can check any value against the video in one click. When a table scrolls on screen, the rows seen at each scroll position are stitched into one table.

What you get

For each video:

• Dataset rows: one row per table row, with the video name, table id, row number, frame time (00:01:07), the cells as an array and as text, the table kind (ruled grid or aligned text), the text-recognition confidence and a link to the frame image.
• Excel workbook (v01_tables.xlsx, one sheet per table) and one CSV per table (v01_table_001.csv, ...).
• HTML report (v01_REPORT.html): each table next to a thumbnail of the frame it came from.
• Frame images (v01_frame_table_001.jpg, ...) and a run summary (OUTPUT: length, tables, rows, processing time, errors per video).

Input

• Video file links: direct links to video files you own or have the right to use (MP4, MOV, MKV, WebM...). Links to video platform pages (YouTube, Vimeo, TikTok...) are refused: this actor does not download from video platforms.
• Or a key-value store in your Apify account holding uploaded video files (for a file on your computer: Storage > Key-value stores > upload, then pick the store here). Optionally list only some record keys.
• Frames sampled per second (default 1): 1 suits slides and screen recordings; raise it for fast scrolling, lower it for long talks. Time and cost grow with it.
• Stitch scrolled tables (default on).
• Screen region (optional): read only part of the frame, e.g. 0,0.1,1,0.8 to skip a menu bar and a webcam strip.
• Maximum video length (default 120 minutes) and maximum file size (default 2000 MB): longer or larger files are skipped.

Example input:

{
  "videoUrls": ["https://example.com/my-screen-recording.mp4"],
  "framesPerSecond": 1,
  "mergeScrolledTables": true
}

Measured accuracy and speed

Measured by us on our own test videos with known contents (2026-09-25). Your videos will differ; treat these as examples, not guarantees.

Test video	Result
54 s screen recording of a spreadsheet, 60 rows + header, 1280x720, scrolled both step by step and continuously (run on Apify, 4096 MB memory)	1 table, 61/61 rows, no duplicate rows, 488/488 cells exactly right; 78 s run time
16 s slide video, 2 table slides + 2 text slides, 720p	35/36 cells exactly right, no false tables
Same slides, 360p with heavy compression	35/36 cells exactly right (36/36 ignoring spaces), no false tables

The scrolling test cost US$0.018 of Apify platform usage at 4096 MB memory, about US$0.02 per minute of video. Most of the time goes into reading each new screen state; a video with fewer changes on screen (slides that stay up for a while) has fewer states to read. We have measured this on one scrolling recording only, so expect the cost per minute to vary with how often the screen changes.

Use with AI agents

Call the actor with videoUrls (a list of direct file links) and read the default dataset: each item has cells (array of strings), frameTime and frameImage, so an agent can quote a value together with where it appeared in the video. The OUTPUT record in the default key-value store summarises the run and lists per-video errors.

Limitations

• Text recognition can be wrong without warning (for example a dropped space or a misread character). Check important values against the linked frame before relying on them.
• Tables are found from ruled grid lines or from text aligned in rows and columns. Charts, handwriting, tables drawn as images inside busy slides and very small text are not handled reliably.
• Tested on English text and digits only. The bundled recognition model (PP-OCRv4) also covers Chinese characters, which we have not tested; other scripts are not supported.
• Use only videos you own or are allowed to process.

Privacy

Your video is downloaded into the run's temporary storage, processed and deleted when the run ends. Results stay in your own Apify storage. Madrasco does not collect or keep your videos or results.

Support

Open an issue in the Issues tab of this actor. Replies are AI-assisted; a human owner can be reached on request.

Built and maintained by Madrasco, with AI assistance. Output is provided as is, without warranty; you are responsible for checking it before use.