Technical and datasets

Sample video frames for machine learning datasets

Create timestamped image samples from a local video for an annotation workflow. Control interval and range, export a ZIP and review dataset quality.

Free to use · No account · Selected video stays local

Video frame workspaceLOCAL PROCESSING

Choose a video from your device

No video upload. Try a compatible MP4, MOV, or WebM file. Up to 2 GB; playback depends on your browser.

The real preview is a 3-second, 640 × 360 derivative of the public-domain FourPeople recording. Published 1280 × 720 trial files use the full source linked in the examples.

The sample opens a small 35 KB video from this website. Your own selected video stays on your device.

Up to 120 captured images / 96 MB per gallery. Sharpness is an estimate; seeking is browser-dependent. The end time is excluded from a batch. Your video stays on your device.

Build an image sample before an annotated dataset

VideoFrame Extractor can help prepare still images for a computer-vision workflow. Choose a local recording, select a meaningful range and sample at a controlled interval. Download the resulting images with timestamped names, then label them in your annotation tool. This workspace does not export YOLO bounding boxes, COCO JSON, class labels or segmentation masks. Those require a separate labeling and validation step.

Sampling decisions affect what the model sees

A one-second interval reduces near-duplicate frames but can miss short events. A smaller interval can show transitions while overrepresenting one scene. Start with a short trial, inspect the images, then decide whether the interval matches the task. Exclude unreadable frames and verify that lighting, backgrounds and object sizes are represented appropriately. Do not assume that a larger number of similar pictures creates a more diverse training set.

Split by source, not only by image name

Adjacent frames from the same recording are strongly related. Randomly placing them across training and evaluation sets can produce an overly optimistic assessment. Track the original recording and requested timestamps outside this tool, and plan splits by an appropriate independent source or scene. Browser timestamp seeking is approximate; use a dedicated decoder if exact frame identity matters. Process short batches within the 120-frame and 96 MB gallery limits, and retain permission to use the footage.

Questions for this workflow

Can I download YOLO or COCO labels?

No. The ZIP contains images and, if enabled, frames.json with filenames, dimensions and player timestamps. It has no bounding boxes or class annotations. Add and validate those separately.

Does the tool automatically remove duplicate frames?

No. Choose an interval and inspect the output. Dataset deduplication and split validation remain separate steps.

Continue your workflow

Read practical guides · Enlarge an image · Report a reproducible problem

Capture uses the shared frame-extractor limits: 2 GiB source, 8,294,400 decoded pixels, 120 pictures and 96 MiB retained images. Start with a short playable sample; codec support and device memory vary.

Preserve the optional ZIP record

Enable the metadata checkbox before Download ZIP to include frames.json. It records source filename, picture names, dimensions, format, player time and capture role. Keep it with your annotation project; it is not COCO JSON and does not prove an exact source-frame identity. Group related recordings before splitting training and evaluation sets.

A collection record you can actually inspect

Executed 25 October 2026 in desktop Chromium. Controlled fixture; results vary with source and browser.

Download the 3-second 160 × 90 H.264/AAC input (35,830 bytes)Download the executed three-image trial · Inspect the CSV · Inspect the JSON

Settings: range 0–2 seconds, count 3; CSV and JSON enabled. The ZIP contains three PNGs at player times 0, about 0.667 and about 1.333 seconds, plus the two logs. It contains no annotations. Keep recording groups independent before labeling or evaluation; a clean capture log does not prevent dataset leakage by itself.

"source","image","player_seconds","width","height","role"
"Sample video · 160 × 90","frame_0001_0-000s.png","0","160","90","frame"
"Sample video · 160 × 90","frame_0002_0-667s.png","0.666666","160","90","frame"
"Sample video · 160 × 90","frame_0003_1-333s.png","1.333333","160","90","frame"

Inspect three real camera captures and their records

On 20 November 2026, the actual dataset workspace loaded our 3-second, 640 × 360 FourPeople preview and captured requested timestamps 0.5, 1.5 and 2.5 seconds. The downloaded ZIP contains three PNGs, captures.csv and frames.json. The preview is a derivative of the public-domain camera recording linked in the source record.

Download the actual captures and records ZIP (1.01 MB) · Download its exact input · Source and experiment record.

To repeat, select “Try real camera preview,” choose timestamps 0.5,1.5,2.5, capture, enable both record exports and download the ZIP. Compare the requested timestamps, dimensions and filenames against the included images. Browser seeking is not proof of an exact original frame index. These records describe captures; they do not label objects, produce YOLO/COCO annotations or decide whether a frame belongs in a training set.