Chocolate-dip chromis · w22_fish3

Chromis dimidiata · Pomacentridae · 92% · high reconstructed model: good
Reviewer: two-tone body, forked tail (39% coral-cut frames)

Where it comes from

Original video (longest visit, 30:40-30:45), the selected fish boxed
Tracking: real frame + crop + 3D render at the estimated pose, species overlay

3D and 4D model

loading 3D gaussians…
Interactive canonical 3D gaussians (drag to orbit, scroll to zoom). This is the rest pose shared by every frame; the 4D part is the per-frame bending / tail motion shown on the right.
4D: canonical model replaying the learned body bending and tail beat
3D turntable of the canonical model

Poster

poster

Species

Chocolate-dip chromis Chromis dimidiata Pomacentridae
92.5%
Axilspot hogfish Bodianus axillaris Labridae
5.9%
Lyretail hogfish Bodianus anthioides Labridae
0.6%
Bluethroat triggerfish Sufflamen albicaudatum Balistidae
0.3%
White-face butterflyfish Chaetodon mesoleucos Chaetodontidae
0.2%

Family: Pomacentridae 93%, Labridae 7%, Balistidae 0% · agreement across crops 97% · 64 crops · bioclip-2.5 (hf-hub:imageomics/bioclip-2.5-vith14, open_clip, zero-shot)

Crops used for the species vote

Data and reconstruction

window22:00 – 32:00
visits11: 25:23-25:26, 26:10-26:12, 26:18-26:20, 26:27-26:29, 26:56-26:58, 27:41-27:43, 27:52-27:54, 28:57-29:00, 30:40-30:44, 31:10-31:13, 31:39-31:41
frames (10 fps) detected / kept232 / 218
valid poses (10 fps)223
frames with background plate219
median size118.0 px
30 fps frames / valid poses422 / 378
method4D GS, fins + bending + tail (30 fps)
model shownv1 (gs_final)
iterations4000
gaussians17568
training time21 min
trainer validation PSNR full / masked27.3 / 18.4 dB
trainer validation IoU vs SAM mask0.821
near-fish band PSNR (fins)27.5 dB
production statusdone
visual note (curation)Chromis dimidiata likely

Diagnostics

Tracklet contact sheet (selection)
Held-out frames: real crop vs 4D render (composited over the plate)
Held-out frames: render on black
Estimated pose: fitted ellipsoid + axes on crops
Frame cleaning: identity / occlusion over time
Frames down-weighted / dropped by the robust trainer

Downloads

point_cloud.ply (canonical 3D gaussians) · poses_refined.json · model.pt (incl. per-frame 4D params) · species.json

← all fish