Heliosoph/da3-base-onnx

Depth Anything 3 Base — Monocular Depth + Camera Intrinsics (ONNX)

Single-view ONNX export of depth-anything/DA3-BASE — the Apache-2.0 any-view model of ByteDance's Depth Anything 3 family (DINOv2 ViT-B backbone, dual-DPT head, camera heads). This cut takes one image per row and emits depth, per-pixel confidence, and a per-image camera-intrinsics estimate.

Two things to know up front:

• Depth is scale-ambiguous (up-to-scale, not meters). For metric depth use Heliosoph/da3metric-large-onnx, or anchor this model's scale against it.
• The pose output is not useful here. DA3 predicts camera pose relative to the other views in the same forward pass; with a single view the extrinsics output is near-identity by construction. For real pose recovery use the multi-view sibling: Heliosoph/da3-base-4view-onnx (same weights, 4-frame window). The intrinsics output is meaningful for single images — a per-image focal-length estimate.

Re-exported from upstream safetensors via the official depth-anything-3 package. Provenance trail: Lin et al. → depth-anything/DA3-BASE safetensors → depth_anything_3.api.DepthAnything3 + thin wrapper → torch.onnx.export → these files. The any-view checkpoints need two exporter workarounds (baked into the script): torch.cartesian_prod (RoPE position grid, no ONNX symbolic) replaced with meshgrid+stack, and the TorchScript-compiled affine_inverse (whose aten::mT is unexportable) rebound to a transpose(-2, -1) equivalent. fp16 sibling via onnxconverter-common with a Cast-node type realignment.

Toolchain: torch 2.4.x (CUDA 12.4), depth-anything-3 0.1.1, opset 17, legacy TorchScript exporter, fp32 trace (upstream bf16 autocast disabled). Conversion script: scripts/export-da3metric.ps1 in the Heliosoph repo. Export validation: fp32 ONNX matches PyTorch to 4.1e-07 max relative error across all four heads; fp16 matches fp32 to ≤4.9e-04; batch>1 verified item-wise against batch=1.

Credit: Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, Bingyi Kang (ByteDance Seed). Paper: "Depth Anything 3: Recovering the Visual Space from Any Views", 2025.

What this repo contains

File	Variant	Size	Use
model.onnx	fp32	~394 MB	Default — matches the PyTorch upstream to ~1e-6.
model_fp16.onnx	fp16	~198 MB	Half precision, I/O stays fp32 (keep_io_types) — drop-in swap.
config.json	—	<1 KB	Upstream DA3 model config (provenance / re-instantiation).

Input / output

	Spec
Input name	image
Input shape	[batch, 3, 504, 504] (NCHW)
Input dtype	float32 (both variants)
Preprocessing	RGB, scale to [0,1], ImageNet mean/std ([0.485, 0.456, 0.406] / [0.229, 0.224, 0.225])
Output depth	[batch, 1, 504, 504] — up-to-scale depth, bigger = farther
Output depth_conf	[batch, 1, 504, 504] — per-pixel confidence
Output extrinsics	[batch, 1, 3, 4] — [R | t]; near-identity for single view (see above)
Output intrinsics	[batch, 1, 3, 3] — estimated K at the 504×504 input grid (principal point at 252, 252); rescale to source dims via K' = diag(W/504, H/504, 1) · K
Dynamic axes	batch only

Resolution is fixed at 504×504 — the ViT position-embedding interpolation bakes the patch count into the trace (inherent to DA3 ONNX exports, not a choice). Resize inputs to match; re-run the conversion script with -Height/-Width for a different fixed resolution (multiples of 14).

How to use

import numpy as np
import onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("model.onnx")

img = Image.open("photo.jpg").convert("RGB")
x = np.asarray(img.resize((504, 504), Image.BILINEAR), dtype=np.float32) / 255.0
x = ((x - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]).transpose(2, 0, 1)[None].astype(np.float32)

depth, conf, _ext, K = sess.run(["depth", "depth_conf", "extrinsics", "intrinsics"], {"image": x})
depth, conf, K = depth[0, 0], conf[0, 0], K[0, 0]

# K is at the 504x504 grid; rescale to the original image if needed:
w, h = img.size
K_src = np.diag([w / 504, h / 504, 1.0]) @ K

License

Apache-2.0 — same as upstream depth-anything/DA3-BASE. (The DA3 any-view Large/Giant checkpoints are CC-BY-NC 4.0 and are not part of this export; Base is the largest permissively-licensed any-view variant.) The ONNX-export step doesn't change licensing — same model, different serialization.