Depth-Anything-1-ONNX

Depth Anything V1 - ONNX (FP16)

FP16 ONNX exports of the two smaller Depth Anything V1 checkpoints, built for Oku3D's depth registry and usable by any ONNX Runtime consumer.

Interface

input   pixel_values     float32  [1, 3, height, width]
output  predicted_depth  float32  [1, 14*(height//14), 14*(width//14)]

height and width are free and take any multiple of the patch size 14; the batch axis is fixed at 1. Feed ImageNet-normalized RGB (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]). The output is relative disparity, near = high, which is the Depth Anything convention and needs no inversion for stereo warping.

Weights are FP16 while the boundary stays FP32, so a consumer hands over what it already has and the casts live inside the graph.

Accuracy

Mean absolute deviation from the PyTorch checkpoint at 518 px over the four example images, as a percentage of the reference map's own depth spread:

Throughput

GeForce RTX 5070 Ti, ONNX Runtime 1.24.4 on DirectML, free dimensions pinned per resolution, fastest of 20 runs on a fresh session (2026-09-21):

Pinning matters: the same graph with free dimensions left open runs on generic kernels and loses a large part of this.

Usage

import numpy as np, onnxruntime as ort

edge = 518
options = ort.SessionOptions()
for name in ("height", "width"):
    options.add_free_dimension_override_by_name(name, edge)

session = ort.InferenceSession(
    "da-v1-small_fp16_opset20_optimized.onnx", options,
    providers=["DmlExecutionProvider", "CPUExecutionProvider"],
)
disparity = session.run(None, {"pixel_values": tensor})[0]  # [1, edge, edge]

License

Apache-2.0, inherited from the checkpoints; see LICENSE. Upstream: Depth Anything, Yang et al., CVPR 2024.