Syncing capture times across five cameras and building per-day sequences in Premiere

I lined up 905 clips from cameras with mismatched clocks to their real capture times using GPS and photos, then automatically built per-day Premiere sequences with a separate track for each camera.

In part 1 I gathered the footage onto one drive. This step lays it out so it can actually be edited. There were five kinds of camera: Osmo Nano, GoPro, Canon 100D, Canon G7X (my wife's camera), and a phone.

Syncing capture times: every camera's clock was different

To edit, you need clips from several cameras placed in the order they were actually shot. But each camera's clock used a different reference. Local time in Tanzania is UTC+3, six hours behind Korea.

The AI pulled the capture time from all 905 files and worked through them camera by camera.

Camera	Clock state	How it was fixed
Osmo (412 clips)	File names carry Korea time	Subtract 9 hours from the time in the file name to get UTC, then add 3 hours for local time
GoPro (215 clips)	A mess	See below
Canon 100D and G7X (276 clips)	Not local time	First wrongly assumed to be UTC, confirmed as Korea time in a later session
Other (2 clips)	Filed separately	Their own track

GoPro

The GoPro asks you to set the time every time you turn it on, and I often punched in something rough just to get past it. Some were probably set to the wrong year entirely, like 2016. So the timestamps on those files couldn't be trusted.

The clue the AI found was the GPS data the GoPro records inside the video. GPS carries accurate time from the satellites. Files with a GPS fix used that time directly. For files without one, it relied on the fact that a wrong clock is wrong by a constant amount. It found ranges of file numbers shot with the same clock setting, worked out the offset for each range, and shifted them back in one go.

File number range	Clock offset
0418–0433	Almost correct
0434–0460	About one day
0461–0479	About 3 hours 20 minutes, in the other direction
0482–0636	About 153 days
0637–0640	About 163 days

That recovered 214 of the 215 clips. The last one (GX010479) got "estimated time" added to its name. Where the GoPro and Osmo were rolling at the same time, comparing the audio waveforms confirmed the sync to within 10–20 seconds.

Canon

For the Canons, I first took them to be on Korea time, then assumed local time and ran two rounds of audio comparison. Both were wasted effort. In the end, a single photo solved it. At the same moment the Osmo showed daytime, the Canon showed us eating dinner. That made it certain it wasn't local time. UTC seemed likely, so I entered +3 hours for now. That got overturned in the next session. Photos stamped 8–10 p.m. by the camera clock were all midday safari scenes (zebras, lions, elephants), and shifting them back 9 hours gave the most overlap with the GoPro GPS and the Osmo. Both the 100D and the G7X were on Korea time. The G7X videos (MP4), though, were recorded correctly and needed no correction.

Per-day sequences in Premiere

Once the times were sorted, the rest was just placing clips by rule. The AI wrote a script that generated sequence files (XML) for Premiere, and I imported them. We tested with DAY 1 alone first, confirmed the approach, and then built everything in one go.

Item	Rule
Sequences	16, one per day, 4K 29.97fps
Day boundary	4 a.m. Anything shot after midnight goes to the previous day
Tracks	V1 Osmo, V2 GoPro, V3 Canon, V4 other. Two audio tracks per camera
Simultaneous shots	Stacked at the same point, one above the other
Between scenes	Gaps compressed to 2 seconds
Markers	Local time at each scene
Clip names	[date time] appended
Scale	2.7K clips scaled up to fit 4K
Left out	Temporary sync files, thumbnail and proxy files, anything under 0.5 seconds, exact duplicates

The 9/26 sequence at around 1 hour 40 minutes. V1 Osmo, V2 GoPro, V3 Canon, V5 photos, V9 phone. Cameras that caught the same moment are stacked on top of each other.

Here it is by day.

Date	Clips	Cameras	Length
9/24	4	Osmo	2 min
9/25	54	Osmo 41, GoPro 13	158 min
9/26	130	Canon 68, Osmo 36, GoPro 26	232 min
9/27	182	Canon 96, Osmo 66, GoPro 20	380 min
9/28	42	Canon 33, GoPro 6, Osmo 2, other 1	91 min
9/29	154	Canon 75, Osmo 44, GoPro 35	454 min
9/30	2	GoPro	2 min
10/1	37	Osmo 21, GoPro 16	387 min
10/2	51	Osmo 36, GoPro 15	367 min
10/3	62	Osmo 47, GoPro 15	281 min
10/4	41	Osmo 31, GoPro 10	310 min
10/5	30	GoPro 23, Osmo 7	322 min
10/6	38	GoPro 27, Osmo 11	129 min
10/7	20	Osmo	244 min
10/8	13	Osmo 9, GoPro 4	198 min
10/9	15	Osmo 13, GoPro 2	119 min

9/29 alone runs over 450 minutes. Opening every clip, laying them out by camera and lining up the times is normally the part of editing I hate most, and it was gone entirely. Once the import finished, Premiere started its own media analysis.

I kept the working files. When more footage turns up, I just run these again.

File	Purpose
clip_times_fixed.csv	Corrected capture times for all 905 clips
make_all.py	Script that builds the per-day sequences. Rerun it when the Canon times change
Handover notes (Markdown)	Locations, per-camera offsets, remaining work, working tips

v2 sequences, rebuilt in the next session

After the Canon times were confirmed, I rebuilt the sequences in a new session.

Task	Result
Canon times	100D and G7X confirmed as Korea time and corrected
Photo track	778 photos (100D 678, G7X 73, GoPro 15, Osmo 12) placed on a photo track as 3-second shots. 343 that exist only as RAW (CR2) can't be read by Premiere, so they were left out for now and listed in a file
v2 sequences	Rebuilt with photos as 17 per-day sequences. Five photos taken in Korea on the morning we flew out created an extra 0923 sequence
New footage	407 clips that had never been backed up, from the G7X, 100D and phone, placed on new tracks, so the existing track layout and the MP3s already exported stayed untouched
Finding missing files	Tracked through gaps in the camera file numbers. The 240 files numbered 4293–4532 on the 100D match exactly one day, DAY 4, so they're probably in the laptop material

The project panel. The per-day MP3 exports at the top, the 17 v2 sequences in 4K 29.97fps below.

In the next post I write up the connection that let AI drive Premiere directly: the Premiere MCP.