Chocolate-dip chromis · w58_fish7

Chromis dimidiata · Pomacentridae · 99% · high reconstructed model: good

Reviewer: two-tone body, forked tail

Where it comes from

Original video (longest visit, 59:01-59:08), 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

Previous version (v1) vs v2

v2 = the same 30 fps data and trainer as v1 (gs_final, 4000 iterations) with 12000 iterations and a gaussian cap of 100,000 — the only change of the quality experiment that passed its review. Other changes (frame selection, water and motion-blur models, white balance, lossless crops) did not.

v1 (previous version): gs_final, 4000 iterations

v2: gs_v2, 12000 iterations

The two turntables play in sync (same orbit and timing).

metric	v1 gs_final	v2 gs_v2	change
iterations	4000	12000
gaussians	18,825	33,531	x1.78
training time (optimisation)	18 min	76 min
held-out frames (within 0.3 s of a training frame)	28 (11)	28 (11)
LPIPS on the fish (lower = closer to the photo)	0.1314	0.1272	-0.0041
PSNR full crop (dB)	26.03	26.34	+0.31
PSNR on the fish mask (dB)	21.08	21.65	+0.57
sharpness ratio render / photo, grain-robust (1 = as sharp as the photo)	0.815	0.838	x1.03
sharpness ratio, raw	0.695	0.755	x1.09
sharp match (photo detail reproduced in place)	0.711	0.770	+0.059
IoU with the SAM mask	0.881	0.877	-0.004

Verdict: better. LPIPS -0.0041, grain-robust sharpness x1.03 (sharper). Differences below 0.0034 LPIPS or x1.12 sharpness are within the run-to-run noise of training.

Held-out frames: fixed per fish (work/quality_exp/<fish>/heldout.json), evaluated identically for both models (test-time fit of pose, colour and bending with the gaussians frozen). Both models were trained on all frames of the same data set, so held-out frames near a training frame favour both equally; this is a like-for-like comparison of the two recipes, not a generalisation test.

Poster

Species

Chocolate-dip chromis Chromis dimidiata Pomacentridae

99.3%

Whitefin damsel Pomacentrus albicaudatus Pomacentridae

0.3%

Lyretail hogfish Bodianus anthioides Labridae

0.1%

Yellowfin damsel Amblyglyphidodon flavilatus Pomacentridae

0.1%

Axilspot hogfish Bodianus axillaris Labridae

0.1%

Family: Pomacentridae 100%, Labridae 0%, Serranidae 0% · agreement across crops 100% · 64 crops · bioclip-2.5 (hf-hub:imageomics/bioclip-2.5-vith14, open_clip, zero-shot)

Data and reconstruction

window	58:00 – 68:00
visits	16: 58:10-58:20, 59:02-59:07, 59:12-59:17, 59:22-59:30, 60:01-60:03, 61:02-61:10, 61:19-61:24, 61:57-62:03, 62:22-62:27, 63:28-63:29, 63:34-63:36, 63:54-63:56, 64:13-64:20, 64:25-64:31, 67:37-67:41, 67:45-67:50
frames (10 fps) detected / kept	684 / 666
valid poses (10 fps)	651
frames with background plate	539
median size	122.0 px
30 fps frames / valid poses	655 / 588
method	4D GS, fins + bending + tail (30 fps)
model shown	v2 (gs_v2, quality experiment rollout)
iterations	12000
gaussians	33531
training time	76 min
trainer validation PSNR full / masked	28.2 / 21.3 dB
trainer validation IoU vs SAM mask	0.885
near-fish band PSNR (fins)	30.5 dB
production status	done
visual note (curation)	Chromis dimidiata likely

Diagnostics

Held-out frames: real crop vs 4D render (composited over the plate)

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 · previous version v1: gs_final point_cloud.ply

← all fish