Depth quality qualitative analysis — v11 step 3000

Date: 2026-07-10 · Model: v11_pushpix step 3000 · Compared against: DA3 teacher

Side-by-side qualitative and quantitative comparison of v11's learned PointmapHead depth output against the DA3 teacher on 4 ARCTIC clips (2 best-hand-pose + 2 worst-hand-pose).

Quick links: Gallery · Metrics · Key finding

Gallery

[**🎞️ Open depth quality gallery →**](../gallery_depth/gallery.html)

4 clips × 4-panel mp4: 1. Raw RGB (input) 2. DA3 GT depth (teacher, turbo colormap 0.2-1.5 m) 3. v11 predicted depth (same range/colormap) 4. Log-ratio bias log(pred/GT) (RdBu_r ±0.7) — red = over-predict, blue = under-predict

Per-clip metrics

Clip	Hand-pose category	AbsRel ↓	δ<1.25 ↑	MPJPE (hand pose)
#7660 s06 espressomachine_use_01	best hand	0.335	0.189	60 mm
#8422 s06 microwave_use_02	best hand	0.327	0.175	61 mm
#8888 s06 phone_grab_01	worst hand	0.336	0.166	540 mm
#14835 s09 espressomachine_use_02	worst hand	0.358	0.173	518 mm

External reference points

DA3 on ARCTIC (in-distribution training set for DA3): AbsRel ~0.15-0.20, δ<1.25 ~0.65-0.80.
v11 pred: AbsRel 0.33, δ<1.25 0.17 → predictions are ~2× worse than DA3 teacher on the same data. Given the small pointmap head (~2 M params) and the SSI loss (which loses metric scale), this is expected.

Key finding — depth is uniform across hand-pose quality

Depth AbsRel is nearly identical (0.33-0.36) across best-hand-pose and worst-hand-pose clips. This means:

The failure mode of the worst hand-pose clips is NOT depth quality — it's specifically the pixel localization (dense root_head soft-argmax fails on occluded/off-frame hands).
Depth head trained uniformly across the val distribution — it doesn't have the bimodal behaviour that hand pose does.
Fixing hand-pose failure modes (presence gating, off-frame handling) would move MPJPE much more than fixing depth.

Ordering of remaining gaps

By expected impact on aggregate abs MPJPE:

Rank	Fix target	Est. gain	Cost
1	Presence gating on lift path	−10 to −20 mm	100 LoC
2	Off-frame clamp_root_to_edge	−5 to −15 mm	50 LoC
3	6D-rot MANO orient (Iter 209 v12)	−5 to −40 mm	200 LoC
4	VGGT point_head reuse (Iter 208 v13)	depth only, marginal on abs	500 LoC
5	Multi-resolution root_head (Iter 209 v14)	pixel −40 to −80 px	300 LoC

Depth head is #4 — modest impact on abs MPJPE. The dominant next moves are 1-3 (hand-pose robustness).

Visual observations from gallery

Looking at panel 4 (log-ratio bias) across clips:

Uniformly slight over-prediction — the mode of log-ratio is +0.1 to +0.3 (pred/GT ≈ 1.1 to 1.4), meaning v11 predicts ~15-30% deeper than DA3.
Depth-normalized structure preserved — panels 3 and 2 show similar gradient patterns (hands closer, background further), just biased upward.
Hand region rarely blue — the hand (typically closest object in ego view) is not systematically under-predicted, which suggests SSI's median normalization is doing its job of decoupling scale.

The uniform bias likely comes from the SSI loss — median-normalization absorbs additive shift, so the model has no metric anchor and drifts freely. Adding a small metric anchor (from GT wrist z on rare valid samples) would fix this.