TITAN TWO

Titan Two and computer vision

Computer vision basics for analyzing gameplay recordings

Computer vision extracts measurements from images or video. For a gaming example, you might count the times a scoreboard appears or measure how long a menu stays open using a recording and a clearly defined visual event.

Define an event you can label by hand

Try detecting when a scoreboard opens in a saved gameplay recording. Define one event as the first frame where the scoreboard title and its panel are both visible. Count the opening once even if the panel remains visible for a few seconds. Include several menus without a scoreboard, as well as a few frames with similar colors or text.

Go through the footage and mark the frame number of each event before trying a detector. Make a second clip for the final check. A separate clip gives you a place to test whether the rule works beyond the examples you used to choose the settings.

Make a read-only comparison

Choose a small region containing the stable title bar and record its bounds and the video resolution. Compare the whole picture with that region. If the game changes its interface scale, verify the region against the new preview before reusing it. OpenCV documents video-file input for a frame-by-frame check, while ConsoleTuner documents the Computer Vision panel in Gtuner IV for viewing supported sources.

Set a matching rule before scoring. For a fixed-rate clip, you might count a prediction as correct if it falls within two frames of a manually marked opening. Match each opening at most once. Mark extra predictions and missed openings separately.

Work through a small validation result

Illustrative data for a clip with 10 manually marked openings. The analysis reports 9 events, of which 8 match the labels and 1 is a false alarm. Two labeled openings were missed. Precision is 8 divided by 9, or about 89 percent. Recall is 8 divided by 10, or 80 percent. These values describe this sample and its matching rule.

Watch the three errors and describe the visible cause if you can identify one. If a similarly colored menu produced the false alarm, retain that menu in the held-out test when comparing a revised approach. The review should end with the annotations and results for the saved clip.

Sources

publishes these guides independently. ConsoleTuner makes the Titan Two device and Gtuner IV.