Ball Toss Analysis

Each bar represents 3 consecutive frames. Height = number of pixels that changed by more than 15 intensity levels. Orange bars exceed the detection threshold.

Each event is a cluster of consecutive high-motion frames — corresponding to a ball being thrown across the table.

Frames extracted at peak motion intensity. The lab was nearly dark — hover to see the brightness-enhanced version.

Frame-to-frame pixel difference, amplified 8× and color-mapped. Bright regions show where the ball (and thrower) moved between consecutive frames.

Extracted from 42–57s in the recovered recording. Raw on the left, brightness-enhanced on the right.

Peak motion = ball closest to camera = too late. The robot needs to detect the ball at release — 1 second earlier, when it's small and far. These early-flight frames are the real training targets.

Each bar shows motion (pixels changed) at a specific time offset from the peak. Negative = before impact. The ball is detectable 1 full second before it arrives.

CLAHE (Adaptive Histogram Equalization) brings out local contrast without blowing highlights. Scroll through each event's trajectory — left = ball far (early), right = ball close (peak).