V1: Initial Distillation Baseline with Relative-Depth Teacher

V1 is the initial distillation configuration. It pairs a MobileNetV3-Small student with a Depth Anything 2 Large (DA2) teacher under standard MSE supervision. Evaluated against metric ground truth, the configuration produces an RMSE of 75.37 m on the NYU Depth V2 validation set — a result attributable to a unit-space mismatch between the teacher output and the evaluation metric, not to optimization or architecture failure.

Configuration

Property Value
Architecture MobileNetV3-Small encoder, single-head depth decoder
Trainable parameters ~3 × 10⁶
Teacher DA2-Large (relative depth)
Loss Mean squared error against teacher output
Training corpus NYU Depth V2
Optimizer AdamW with cosine annealing
Hardware NVIDIA L40S 48 GB (NYU Greene HPC)
Status Baseline; superseded by V3

Training Target Mismatch

DA2-Large produces relative depth: per-image depth ordering normalized to the unit interval [0, 1] without an absolute metric scale. The training procedure minimized MSE between the student prediction and these normalized targets. The student converged successfully against the teacher within the relative-depth space.

The evaluation metric, in contrast, was computed in metric units (meters) against the NYU Depth V2 ground truth, which spans approximately [0.1, 6.0] m for indoor scenes. Predictions in the [0, 1] range compared against targets in the [0.1, 6.0] range produced the observed 75.37 m RMSE.

flowchart LR RGB["RGB frame
NYU Depth V2"] --> DA2["DA2-Large teacher
(relative depth)"] DA2 --> RELDEPTH["Predicted depth
values ∈ [0, 1]
(normalized)"] RGB -.-> STUDENT["MobileNetV3-Small student"] STUDENT --> SPRED["Student prediction
values ∈ [0, 1]"] RELDEPTH --> LOSS["MSE loss"] SPRED --> LOSS LOSS --> OPT["AdamW optimizer
(unaware of unit space)"] OPT -.-> STUDENT GT["NYU ground truth
values ∈ [0.1, 6.0] m
(metric)"] --> EVAL["Evaluation metric
RMSE in meters"] SPRED --> EVAL EVAL --> RESULT["RMSE = 75.37 m"] style DA2 fill:#fde2e2 style RELDEPTH fill:#fde2e2 style GT fill:#e8f0ff style EVAL fill:#fff3cd style RESULT fill:#fde2e2

Diagram source: assets/diagrams/models/v1-unit-mismatch.mmd — refine in mermaid.live and export as v1-unit-mismatch.svg to override the embedded version.

The configuration therefore exhibits consistent internal optimization but inconsistent alignment between the supervision signal and the evaluation metric. The model learned the teacher’s depth ordering correctly; the ordering had no metric anchor.

Identification Procedure

The mismatch was identified during post-hoc evaluation of training output statistics. A single-line verification step in the dataloader at training initialization would have surfaced the inconsistency:

# Verification at initialization
print(target.min().item(), target.max().item(), target.mean().item())
# DA2 output:        ~0.0,  ~1.0,  ~0.5         (normalized)
# NYU GT (m):        ~0.1,  ~6.0,  ~2.5         (metric)

Subsequent configurations include input-space verification at training initialization as a standard pre-flight check.

Resolution Path

Three independent modifications were required to produce the first working configuration (V3). V1 had only the first issue; V2 confirmed the latter two were also limiting:

  1. Metric-scale teacher. DA2-Large was replaced by DA3-Metric-Large, which produces depth in absolute meters by construction.
  2. Loss formulation. MSE was replaced by berHu loss, which transitions from L1 to L2 at the 80th percentile of per-batch residual magnitude. This bounds the gradient contribution of far-field pixels where teacher uncertainty is highest.
  3. Multi-task balancing. Kendall uncertainty weighting was added to balance the depth and segmentation losses without requiring a manual weight sweep.

V1 contained issue (1). Issues (2) and (3) became visible only after issue (1) was resolved, since unit-mismatched supervision dominates any signal that loss reweighting could provide.

Reproducibility and Provenance

The V1 configuration is preserved under archive/v1-v3-baseline/ and is reproducible via the v1-v3-baseline git tag. The MobileNetV3-Small encoder interface was maintained through V3, then deprecated at V4 when the architecture transitioned to EfficientViT-B1.

Findings

The V1 configuration establishes a foundational requirement of the distillation pipeline: the teacher’s output space must be expressed in the units used for evaluation. The 75.37 m RMSE reflects unit-space inconsistency, not a deficiency in optimization, architecture capacity, or training duration. This result motivated the input-space verification practice adopted in all subsequent configurations.

→ V2 reports the diagnostic study that ruled out loss-weighting as the dominant cause of the V1 result. → V3 reports the first configuration to produce metric-scale predictions.