Duplicate frames: what two identical captures and one nearby frame showed
The same camera moment exported twice produced the same PNG hash in this trial. A capture 0.1 second later had a different hash. Exact-file hashing found the repeat; it did not tell us whether the nearby image was useful for our task.
A real-camera test, not a generated slide
We used Xiph's Johnny camera sequence, explicitly listed as public domain. The downloaded WebM was 6,654,698 bytes, 1280 × 720 and reported 10.017 seconds in Chromium 151.0.7922.34. On 20 November we captured PNG at requested 0.5 seconds twice, then at 0.6 seconds through the same workspace.
| Capture | Bytes | SHA-256 prefix |
|---|---|---|
| 0.5 s, first | 908,671 | 213d6a731cb2… |
| 0.5 s, second | 908,671 | 213d6a731cb2… |
| 0.6 s | 918,629 | cb993c954f51… |


Choose what counts as repetition before deleting anything
For a slide summary, two pictures showing the same complete slide may be redundant even if the pointer moved. For expression or motion analysis, a small change can be the reason to keep both. In a dataset, nearby pictures from one recording may belong in the same split rather than being treated as independent evidence.
This tool does not automatically find or remove duplicates. Use a small interval batch as an overview, select two pictures for the built-in comparison and remove a picture only after inspecting it. Keep the original ZIP while reviewing. In another codec, two semantically identical moments may produce different decoded pixels or file hashes.
Repeat and keep the capture record
- Download the public source WebM and open it in the extractor.
- Capture at 0.5 seconds twice without moving the position, then at 0.6 seconds. Save each PNG before comparing its hash.
- For a dataset sample, use the dataset workspace. It can include captures.csv and frames.json in the ZIP. Those records document the requested captures; they do not contain YOLO boxes, COCO annotations or validated source-frame indexes.
- Record why each picture was kept and group related scenes before training/validation splits.
Download the repeated capture · Full hashes and environment. The previous version of this guide incorrectly said CSV export was unavailable; this revision corrects that statement.