PixelOkay
● Video stays on your device
Tools
PIXELOKAY VIDEO LAB · AI-ASSISTED INSPECTION

Activity Review Studio

Turn video movement into a reviewable, timestamped activity report. Scan visible motion offline, or apply a free pose model to estimate supported human postures.

◉ Timestamped findings⌁ Optional on-device pose AI✓ Human review⇩ CSV · JSON · SRT
ANALYSIS WORKSPACE

Find the action. Verify the moment.

All video frames are analyzed locally. The optional AI model is downloaded only when you choose pose analysis.

Video inspection LOCAL

Review highlighted samples and activities on the timeline.

No video loaded
⌁ ◉ ✦

Discover the moments that matter.

Bring your own video or open the sample to explore timestamped motion findings.

Activity timeline

Ready to scan
Low motion / standing Activity Review suggested

Timestamped activity summary

Select a finding to seek directly to its first frame and edit its label.

0 findings
0Frames sampled
0Segments
0Need review
0%Video reviewed
Analyze a video to see an editable timestamped summary here.
Import a video or try the sample to get started.

Useful for editors, researchers and creators

01

Locate moments quickly

Jump from a report entry straight to its corresponding video timestamp.

02

Human-in-the-loop labels

Review estimates, split or correct segments, and preserve manual notes.

03

Private by default

Frame analysis runs inside your browser; no video is uploaded to PixelOkay.

Frequently asked questions

Is this truly activity recognition?

Pose mode uses a free on-device human pose model and rule-based posture estimates. Offline motion mode measures visual change only; it cannot tell which activity caused it.

Can it recognize every action?

No. It supports a limited set of visible single-person poses and broad motion categories. Complex interactions, objects, sports and precise action names need human verification.

Does my video leave the browser?

Video frames are processed locally. Optional pose mode downloads model code and weights from third-party hosts; videos are not intentionally uploaded.

How accurate are timestamps?

They reflect sampled video times rather than frame-perfect starts and ends. Increase sampling detail for more precise boundaries.

Using Activity Review Studio

Review sampled video motion or optional posture cues as editable, timestamped intervals. Use the analysis to locate footage that needs attention, then correct labels before sharing a report.

  1. Load a video or the sample and choose visual-motion review or optional pose-based posture mode.
  2. Set a bounded analysis range, run the analysis and inspect each suggested interval alongside its sampled frames.
  3. Correct labels and times around cuts or camera movement, mark reviewed segments, then export CSV, captions, text or project JSON.

Example: On a short sports clip, use high-motion intervals to find likely action sections, then check the surrounding frames and correct segments where a camera pan or cut caused a misleading cue.

Limits and privacy: The tool does not classify arbitrary actions: motion can come from camera movement, exposure changes or cuts, while pose labels and confidence are heuristic. Analysis is limited to 180 seconds and 240 sampled frames; timestamps are approximate. Optional MediaPipe and local text-model setup retrieves external runtime/model files and may disclose the visitor IP. Project JSON omits the source footage. The optional MediaPipe pose path remains unverified after the pinned dependency problem; visual-motion review is a separate mode.

Related tools

LOCAL AI SETUP

Set up AI on this device

First, a model needs to be loaded on your device

Activity Review Studio offers optional WebLLM text summaries of timestamped results. It does NOT directly classify video frames. Offline motion works without downloads; MediaPipe pose mode separately downloads its own model only after you click Analyze. WebLLM downloads start only on your explicit Download & Load Model action.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Set up assistant
  4. 4. Summarize results

Device & model resources

Measurements may be approximate
Device RAM / model memoryChecking device information…

Live available RAM and GPU memory are not exposed by most browsers.

Browser website storageChecking site storage…

Site-wide approximate origin quota, shared with other PixelOkay tools. This is not RAM.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this website origin for WebLLM and Transformers.js model caches, plus recognizable model files in other cache buckets. This includes caches from other PixelOkay tools, not only this tool. No site data or browser databases are cleared automatically.

Run a scan to inspect recognized model caches before optional cleanup.

Unknown IndexedDB, OPFS and browser-managed models remain untouched. Other AI tabs can hold models in use; close them before cleanup.

Having trouble downloading or loading AI?

Refresh storage/cache, retry, select a smaller compatible model, or inspect the recognized caches. Storage deletion is optional and does not fix network/GPU problems.

Clearing all website data can erase saved work across PixelOkay; this tool never performs that action.

Check compatibility or select Download & Load Model to begin.

Technical details and safe recovery

WebLLM 0.2.87 · selected model. No private video frames in diagnostics.

Do not clear all browser website data without exporting your projects and files. This tool does not access or remove them.

The model analyzes only the bounded, redacted timestamped segment summaries on your device. AI suggestions are not proof of reproduction or a confirmed root cause.

CROSS-TOOL MODEL CACHE CLEANUP

Delete recognized AI model downloads?

This may make other PixelOkay AI tools redownload shared model weights. Saved projects, imported files, and browser databases are not targeted. Unknown caches are skipped. Close other AI tabs first.