Same Video File, Same Threshold, Three Answers: Caption Safe-Zone Frames Are Matched by Grid, Not by Count

说明与用途

由于采样帧对齐错误导致字幕越界警报误报。通过在采样前将两段视频重新对齐到帧索引网格，消除了误报，同时保持了有效检测。

问题

A gate that compares burned‑in captions by diffing two renders reported false violations due to misaligned sampling grids.

结果

Aligning samples by frame index rather than computed timestamps removed false caption‑overlap alerts without affecting legitimate detections.

工作流

• Render preview.mp4 (no captions) and final.mp4 (captions burned in).
• Measure clip duration and compute sampling interval (e.g., 0.736117 s).
• Sample 60 frames at equal intervals from each file using ffmpeg’s fps filter.
• Subtract corresponding samples to isolate caption pixels.
• Check if caption pixels cross the safe zone; flag violations.
• Detect misalignment by ensuring both files share the same frame count, hash, and median difference.
• If misalignment suspected, re‑time both files with setpts=N/FRAME_RATE/TB before fps sampling.
• Re‑run the gate to confirm no false positives and unchanged valid detections.

证据

The gate derives its sampling interval from each clip's actual measured duration. For this pair, the computed value was 0.736117 seconds.

来源

Re‑timing both files onto a frame‑index grid before sampling eliminated the misalignment.

来源