VIDEO DEMO / CNN DETECTION + APPEARANCE TRACKING
Track it. Remix it.
A clear single-ball clip, a pretrained CNN, and a little movie magic. Turn the tracked ball into a capture ball, a glowing orb or your own artwork.
Choose your video
Soccer · one clear ball
Choose a clip above, then Analyze full video. CNN samples cover the entire clip at 10 frames per second, including its final frame. Nothing is uploaded.
Inspect detector output (after suppression)
These saved box/class/score records have already passed the detector wrapper’s score filtering and non-maximum suppression. The app threshold filters them again; lowering it cannot restore a box suppressed upstream.
Analyze the clip first.
Ready to analyze the full video.
Original + model evidence
Your remix
No CNN results yet.
Input frames are centered in a padded square to preserve ball proportions during CNN resizing. Yellow outlines mark CNN positions; blue outlines mark appearance-tracked positions. Replacement size changes the artwork only. The CNN labels “sports ball”; it does not distinguish soccer from every other ball. Tracking begins at the largest detected ball, then prefers the nearest candidate. For a sufficiently colorful CNN-detected ball, an appearance tracker measures matching colored regions in the decoded frames when the CNN misses. It does not invent a CNN detection. If neither method supplies a position, the original stays visible. Between valid samples, positions are interpolated—those positions are not new CNN detections. Replacement is a 2D overlay, so it can cover a foot during occlusion.
MEASURE IT / PORTFOLIO EVIDENCE
Does the effect follow the right ball?
Pause, choose a sampled time, then enable labeling and click the center of the intended ball in the original view. Label a variety of frames, including fast motion and occlusion. Click “Ball absent” when it is out of view. Your labels are separate from model predictions.
No manually labeled samples yet.
Center error is normalized by the image diagonal and applies only to labeled visible-ball frames with a detection. Report misses and false detections on absent frames too. Detection coverage alone is not accuracy.
Compare the ball videos
Choose a video at the top. The longer juggling clip adds kicks and changing ball height; basketball is a harder dribbling test with many misses. Select Ping-pong, then Analyze full video. The orange ball moves quickly and sometimes leaves the frame. The CNN can miss it; appearance tracking follows matching visible pixels after a CNN anchor. Counts and colors distinguish the two sources. When the ball is out of view, no replacement is drawn. Clips are local, and each switch clears old results and labels.
What you can make your own
Compare confidence settings, improve tracking through fast motion, design a new effect, or evaluate on a second clip. Build a report with annotated errors and a short before/after demo. Explore how the tracker works · Build the Python video project →
Starter clip: Soccer Ball Rolling Across Turf During Training, Mixkit contributor, Mixkit Stock Video Free License. About seven seconds, one clear ball; resized and audio removed. Keep the original license when sharing. Ping-pong: Ping Pong Paddle Striking Ball in Mid-Game Action, Mixkit contributor, under the same Mixkit Stock Video Free License. All clips were resized and their audio removed. Longer soccer: Skilled footballer demonstrating ball juggling prowess, Mixkit, under the same free license. Basketball: Basketball player dribbling ball, Marco Lopez / Mixkit, under the same free license.