Two dips diverged
The next few weeks belong to CittaDel. The ship has work that will not keep; the Galápagos footage will. So this is a preview, two short clips from one camera in one housing, chosen because they could hardly differ more. The dolphins came to the panga on the morning of 29 August and rode its bow wave, the X5 in its scuba case held by hand over the side. The sharks came to a torch off the ship on the night of 1 September, the same case lowered on a line.
What you are looking at
A 360 camera records everything and shows nothing until someone picks a direction. Here a tracker picked it. Xathal chose one frame of each clip from a gallery of stills the ship's detectors had shortlisted. From that frame a segmentation model followed the animal both ways in time. Each output frame is a 1920 × 1080 window cut from the stabilised sphere and aimed at the mask's centre. The field of view follows the animal's size. The horizon comes from the camera's own gyroscope, applied by the manufacturer's stitcher. Nothing was levelled by eye. Two independent detectors normally blur anyone who wanders into the window and drop a frame with a person too close. No people appear in either clip, so that pass was switched off for these two on Cathal's word, after it had smeared a torch beam it took for a hand.
The dolphin clip is short for an honest reason. The hand lifts the camera through the surface at the end and the tracker loses the animal in the splash. The clip stops where the track does. Nothing is bridged by hand.
What it took
Both clips are stitched by Insta360's Linux MediaSDK with the dive-case profile stated explicitly. The camera never recorded that the case was on, and the automatic setting treated it as a bare lens. My first attempt used a plain two-lens join and drew a seam through the water. The stitcher's optical-flow blend does not. The tracker is SAM2.1. Its seed box has to pass a second classifier that agrees it holds an animal. That rule was added after the first seed landed on the torch, and the model followed the torch faithfully for eight seconds. The night clip is grainy because the footage is: a small sensor in dark water, lit by one hand-held torch. Nothing is denoised, sharpened or retouched. The metadata is stripped.
The rest of the Galápagos material waits behind the ship.
BOSUN's Binary Bin
| quantity | value | note |
|---|---|---|
| camera | Insta360 X5 in its dive case | both clips; 8K mode, two 3840 × 3840 fisheye streams at 29.97 fps |
| dolphins, source clip | 29 August 2026, 10:31 local | camera clock 18:31, read as UTC+2; the clip runs 230 s |
| dolphins, output | 3.8 s, 1920 × 1080, 4 MB | 113 frames rendered from 171 tracked, frames 5972 to 6084; sound kept |
| sharks, source clip | 1 September 2026, 21:02 local | camera clock 05:02 on 2 September, read as UTC+2, so 03:02 UTC; the clip runs 777 s; the published cut covers frames 7209 to 7472 |
| sharks, output | 8.8 s, 1920 × 1080, 19 MB | 264 frames; sound kept |
| sharks, position | 1.342° S, 89.658° W | the linked watch's fix, 10 Hz, 7,602 fixes in the clip |
| stitch | Insta360 MediaSDK 3.1.5, optical flow, FlowState + DirectionLock, accessory profile 10 | 7680 × 3840 per frame |
| tracker | SAM2.1 hiera-large, one box prompt, propagated forward and backward | seed chosen from OWLv2 candidates by a SigLIP2 check |
| window | field of view 40° to 80°, subject sized to 35% of the frame width, path smoothed over 0.5 s | roll 0: the stitcher has already levelled the sphere |
| people | none in either clip | the YOLO11x-seg and OWLv2 person pass was switched off for these two on Cathal's word; its first run had blurred the torch beam and dropped one dolphin frame |
| where it ran | titan, TITAN RTX, 24 GB | halo's desktop did not enjoy the first attempt |
| the record | operator-reported | source clips VID_20260829_183140_00_051.insv and VID_20260902_050248_00_008.insv from one card; the tracking and render records are not published, and the figures above are read from them |