wrote a Blender script to auto-trim dead frames from mocap takes — tired of finding where the action starts by hand

The per-actor warmup data pays off way later than when you collect it, which makes it easy to skip at the time. We ended up keeping a small JSON file per actor with their typical warmup range from the first session. By session three or four the auto-trim was basically self-tuned to each person without us having to touch it. Just a lookup at trim time and a slight offset to the motion threshold start frame. Makes retakes on later sessions go noticeably faster. You're not babysitting the trim window on every single import.

yeah this broke my script on literally the first real batch i ran it on lmao. one performer could not stop settling, small hip shifts, weight transfers, nothing that looked like action but enough to trip a naive velocity threshold constantly.

fixed it by requiring N consecutive frames above threshold rather than any single frame crossing it:

def find_motion_start(fcurves, threshold=0.02, min_run=8):
    start = int(min(fc.range()[0] for fc in fcurves))
    end = int(max(fc.range()[1] for fc in fcurves))
    for f in range(start, end):
        consecutive = 0
        for check_f in range(f, min(f + min_run, end)):
            vels = [abs(fc.evaluate(check_f + 1) - fc.evaluate(check_f)) for fc in fcurves]
            if max(vels) > threshold:
                consecutive += 1
        if consecutive >= min_run:
            return f
    return start

min_run=8 basically eliminates false positives from fidgeting in my sessions. might need tuning depending on capture rate.

The asymmetry makes sense if you think about it in terms of energy state transitions. "Started moving" is a transition into a high-energy state: fast, well-defined, easy to detect. "Stopped moving" is a transition out of one, and actors almost never do that cleanly. They relax, breathe, shift weight, glance at the director. None of that is the captured action, but it's not stillness either.

What worked better for me than a raw threshold crossing: require a sustained window of consecutive frames below threshold before committing to an end point, around 12–15 frames at 30fps, roughly half a second. Single-frame dips get ignored entirely. Only once you've had a genuine lull does the script mark a candidate end, and then you walk back through the window to find the last frame with real motion above it. Two-pass, but it catches most of the "actor relaxed but didn't freeze" situations cleanly.

The failure mode that still beats it: an actor holding a final pose actively, like sustaining a combat stance at the end of a fight take. They're never truly below threshold so the window never fires. For those I gave up on auto-detection and just flag the take for manual review. Better than a wrong trim point silently corrupting a batch.

yeah logging the range is essential for batch work. i also added a per-take confidence score to the output, just the ratio of frames above the motion threshold to total usable frames. low-confidence takes get flagged so i know to double-check them:

confidence = active_frames / usable_frames
row = [take_name, start, end, usable_frames, f"{confidence:.2f}"]

useful for catching takes where the actor half-moved at the start and tripped the threshold before the actual action began. without that flag those were silently getting clean-looking timecodes when they really needed a second look.

same problem, possibly the same actor lol. fix that worked for me: instead of triggering on any single frame above threshold, require some minimum number of consecutive frames, 8 to 10 at 30fps. a quick weight shift or nervous head turn almost never sustains that long. actual intentional motion almost always does.

doesn't kill all false positives but cuts them down a lot.

The CSV format is worth doing even for small batches. Once you have start frame, end frame, and usable duration as structured data, patterns start appearing. One actor always needs 30+ frames to settle, a specific session consistently has long wind-downs, certain takes have unusually small usable windows. That kind of data starts feeding back into how you give direction during capture, not just how you clean up afterward.

Worth logging the session name and take filename alongside the trim data if your naming convention is consistent. Turns the whole thing into an actual production record rather than a one-time cleanup log you'll never look at again.

also worth logging which threshold value was active when the trim ran, especially if you're using per-session or per-actor overrides. two months later when you're re-processing takes and the trim points don't match your old logs, you'll want to know what parameters generated them. found this out when a session got re-exported from the suit software with slightly different curve smoothing and all my old trim frames shifted by 5–10, making the original logs look wrong when they weren't.

the per-actor warmup pattern thing is genuinely useful data. ran the script on a full session once and noticed one actor was consistently sitting at 50+ frames before any meaningful motion. turned out they had a habit of slowly settling into their start pose rather than hitting it and holding still. once we knew that we briefed them before takes and their warmup dropped by like half. nice to have something concrete to point to instead of just vibes about why setup takes forever.