Real video · Pre-analyzed AI

Video Volume

See motion as a shape. Explore real AI masks across time, then pull out any sampled moment.

8 real clips · 36 analyzed frames each · No model download

Loading analyzed clip…

Loading analyzed clip…

The selected moment

The bright plane is this real video frame, with its AI mask overlaid.

    Explore space and time

    Rotate the volume to follow the shapes each object leaves through time.

    Drag the volume to rotate it. Use the timeline for exact frame selection, or choose Drag time to scrub directly on the canvas.

    Fullscreen gives the volume more room. Scroll or pinch to zoom, or use the zoom buttons and camera sliders. Reset view restores the camera without changing the selected moment.

    Every contour is a real model prediction at one sampled moment. Missing detections remain gaps; object IDs are associated by box overlap.

    Source excerpt

    Original frame rate, without masks. The volume above uses 36 samples from this three-second excerpt.

    Cat relaxing on a sofa.webm by Ivan Kurmanov. CC BY 2.0.

    Three-second excerpt; audio removed; resized to 640 × 360. Frames sampled at 12 fps for AI analysis. Excerpt starts at 12 seconds in the original. Clip, poster and frame atlas retain the source license.

    How the analysis works

    These are real model predictions computed in advance, so switching clips needs no AI processing in your browser. DETR detects known object classes; SAM 2.1 segments each detected box. Every colored contour comes from a predicted pixel mask.

    Object IDs use class-matched box overlap (IoU), not SAM’s temporal memory. Fast motion, occlusion and missed detections can split or swap IDs. Filters select the analyzed classes; this is not open-vocabulary text prompting.

    Models, sampling and downloadable results

    DETR ResNet-50 (q8), confidence ≥ 0.70, duplicate-box suppression at IoU 0.60, up to 3 detections per class and 4 per frame. SAM 2.1 Hiera Tiny (uint8), highest predicted-IoU mask, logits > 0, 256 × 256 masks. Association IoU ≥ 0.15 with two missed-frame grace (up to 0.25 seconds between detections). Twelve samples per second; no interpolated masks.

    Detector · Segmenter · This clip’s results and pinned model revisions (JSON)

    Eight openly licensed real videos, analyzed offline on CPU. Playback starts only when you press Play. No uploads or paid APIs. Confidence scores describe model predictions, not verified accuracy.