README.md

Sample data

Six cases the offline inference scripts run on out of the box — three point-force and three wind-force stills, at 832×480. They are the same images the interactive demo offers as gallery presets (demo/assets/), with the same force values, so a case looks the same whether you reach it through the demo or through inference*.py.

images/          point_1..3.png, wind_1..3.png   (832x480)
point_force.csv         3 rows   a single push
wind_force.csv          3 rows   a single global wind
point_force_change.csv  3 rows   the push reverses halfway
wind_force_change.csv   3 rows   the wind reverses halfway

Pick one with --force_type:

python inference_causal_rolling_forcing.py \
    --config_path configs/dmd_everything.yaml \
    --checkpoint_path <PATH_TO_DISTILLED_STUDENT_CKPT> \
    --force_type point_force \
    --output_folder outputs/

What the columns mean

Forces are 0..1 magnitudes and degrees, matching what the demo's drag produces. The inference scripts pin min_force=0.0, max_force=1.0, so the numbers are used as written rather than renormalised against the file.

Angles are counter-clockwise from +x. Point-force anchors are pixel coordinates with y measured up from the bottom, which is the convention the dataset expects (coordy / height then flipped in the signal builder). The demo's anchor_y is measured down from the top, so coordy = height - anchor_y.

file	columns
point_force.csv	image, angle, force, coordx, coordy, width, height, caption
wind_force.csv	image, wind_angle, wind_speed, width, height, caption
point_force_change.csv	image, angle1, force1, coordx1, coordy1, angle2, force2, coordx2, coordy2, width, height, change_at, caption
wind_force_change.csv	image, width, height, change_at, caption, wind_speed_1, wind_angle_1, wind_speed_2, wind_angle_2

change_at is a fraction of the clip (0.5 = halfway). Drop the column and the change point is sampled instead.

Adding your own

Drop a 832×480 PNG in images/ and add a row naming it. The loader globs images/*.png and keeps only rows whose image matches a file present, so the four CSVs can share one folder and a half-finished row is simply skipped.

The two *_change.csv files reuse the same six images: the second force is the first one reversed (angle + 180), which is the mid-clip reversal the README describes. Change the *2 columns to make it something else.

Note

The paper's numbers come from larger benchmark sets that are not distributed here. These six cases are for checking that a checkpoint runs and behaves sensibly, not for reproducing quantitative results.