Input

Pipeline

Visualization mode
4 64
16 16384
Patch size (px)
GOP (group of pictures)
Splits sampled frames into GOP groups. Each group produces exactly one GOP canvas: the group's first frame stays whole as the I-frame, and each later frame gets its own P section below it. So GOP=4 means each group is I P P P. For fixed GOP=N, the number of packed canvases is ceil(sampled_frames / N). Example: 32 sampled frames with GOP=4 gives 8 canvases; 32 sampled frames with GOP=8 gives 4. Codec-stream mode adaptively groups by saliency energy, targeting roughly 8-64 sampled frames per group.

Patch selection visualization

Cumulative patches over time

Indigo: codec method — selects patches within frames according to saliency, so the curve rises in bursts. Cyan (dashed): uniform full-frame sampling — evenly samples complete frames from the same time window, targeting the same sampled-frame count as codec when the budget allows. Each step costs patch_size^2 patches, regardless of the preview frame resolution. The dotted line marks the requested budget.

Packed canvases (one per GOP group)

Each canvas is one GOP group rendered in I/P structure: the group's first frame is the I-frame kept whole on top, and each later frame gets its own packed P-frame section below in time order. Fixed GOP=N means one canvas per N sampled frames.

Raw JSON

Full reproducible record of this run (params, frame ids, group spans). Collapsed by default — click to expand.