Pi4 / examples /README.md
yyfz233's picture
Set track trails to 1.5 seconds and remove three examples
1037fac
|
Raw History Blame Contribute Delete
4.14 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade

Space Example Videos

These are RGB input videos, not precomputed model outputs. The Space bundles byte-for-byte copies of the main repository's three updated examples/ videos and three compressed RGB input videos from the project website. The skating clip uses the generic figure_skating.mp4 filename. Selecting a cover loads the clip and its sampling preset; it does not run inference until Reconstruct & Track is clicked.

Video Stored size Source frames Space stride Sampled frames Coverage
figure_skating.mp4 512×288 750 at 12.5 fps 1 150 First approximately 12 seconds
kart.mp4 512×288 300 at 15000/1001 fps 2 150 Full approximately 20-second clip
cycling.mp4 288×512 356 at 30 fps 2 150 First approximately 10 seconds
modern_home.mp4 216×384 722 at 30 fps 5 145 Full approximately 24-second clip
waterfall_surfing.mp4 384×216 751 at 30 fps 5 150 Approximately 25 seconds, through frame 745
rooftop_pursuit.mp4 384×156 837 at 12000/1001 fps 2 150 First approximately 25 seconds

Figure Skating uses source frames 0–149. Kart and Cycling use source frames 0, 2, ..., 298. Frame limits count frames after sampling. All example presets enable tracking with query frame 0, and all sampled frames run in one joint forward pass at a fixed 512-pixel longest image side. The website clips retain their existing compressed dimensions and are resized by the loader: landscape inputs become 512×288, portrait inputs 288×512, and Rooftop Pursuit 512×208. Upscaling does not add source detail. The presets favor denser sampling for fast action, while the slower home clip covers its full duration within the same 150-frame cap. These are memory-bounded live inference presets, not the website's full-length results.

The repository's full Cycling example uses 178 frames at stride 2, covering source frames 0, 2, ..., 354 (the full approximately 11.87-second clip). The Space intentionally retains its 150-frame cap to bound GPU memory use; compression does not discard the remaining source frames from the input file.

Encoding

The original three MP4s use H.264 (libx264, CRF 20, slow preset), YUV420p, and video-only streams with fast-start metadata for browser playback. The three website RGB inputs are already compressed and are copied without another lossy encode; together they add about 2.8 MB. Every source frame and its timing are retained: stride is applied by the inference loader, not baked into the video. All MP4s are tracked with Git LFS.

The skating clip uses the generic name figure_skating.mp4. Kart is encoded at 512×288 rather than the old 384×216; this matches the previous inference resize and does not add source detail. Cycling is stored at 288×512 rather than the Space's previous 432×768 encoding or the original 720×1280 clip. These are lossy encodings, so predictions need not be numerically identical to runs on the previous files.

Covers and Attribution

Regenerate the embedded JPEG covers from the bundled videos:

python scripts/build_example_covers.py > example-covers.json

catalog.json is the shared source of titles, filenames, strides, card metadata, and cover-frame indices for the app and generator. Covers use frame 30, except Rooftop Pursuit, which uses frame 210 to show the action rather than an empty doorway. Padding preserves the aspect ratio of portrait and ultrawide clips.

The three added inputs come from these paths relative to the project website:

Space video Website RGB input
modern_home.mp4 assets/demos/luxhome/input.mp4
waterfall_surfing.mp4 assets/demos/waterfall1/input.mp4
rooftop_pursuit.mp4 assets/demos/007/input.mp4

Figure Skating and Kart also derive from the paper website's RGB input assets; Cycling is the supplied third-party cycling clip. Original rights remain with the respective owners. No website point clouds, tracks, or rendered result videos are bundled. All six examples run live inference after submission.