depth-heads

Depth Heads

A systematic study of monocular depth head architectures operating on frozen vision transformer features. Given a dense spatial feature grid from any frozen backbone, what is the most parameter-efficient architecture for per-pixel metric depth prediction?

Standard practice treats the backbone and depth decoder as a joint system. Recent universal encoders produce spatial features of sufficient quality that the backbone can remain frozen while a lightweight head is trained on depth data. Under this regime, the head is the only variable.

This repository contains an arena framework for rapid comparison of depth head candidates and a collection of architectures spanning conventional decoders through novel minimal-parameter designs. All heads consume the same spatial feature tensor and produce per-pixel depth maps. The reference backbone is EUPE-ViT-B (86M parameters, frozen), but the framework is backbone-agnostic — the same heads can be evaluated against any frozen ViT that produces a stride-16 spatial feature grid.

Heads

Nine architectures, all consuming a [B, 768, H, W] spatial feature tensor and producing [B, 1, H_out, W_out] metric depth in meters over the 0.001–10 range. Each head lives in its own folder under heads/ with a single head.py implementation.

Arena Framework

arena.py runs any head by name against cached NYU Depth V2 backbone features. The arena pre-extracts features once, then each candidate trains and evaluates without touching the backbone again. Training is SILog loss at 416×416 resolution against the indoor depth label space; evaluation reports root mean squared error (RMSE) on the test split.

Status

Heads are implemented and importable through the heads/ registry. The arena screening sweep across all 9 heads has not yet been run on a fresh NYU Depth V2 cache; results will be published here when available.