dinov3

kornia/dinov3

ONNX export of DINOv3 ViT-S/16 for vision-rt, producing a global image descriptor (the CLS token) for retrieval, visual place recognition and scene-change detection on NVIDIA Jetson.

All model credit belongs to Meta AI (Siméoni et al.). This repository contains only a format conversion of facebook/dinov3-vits16-pretrain-lvd1689m — no retraining, no fine-tuning, no architectural change. Use of these files is governed by the DINOv3 License, reproduced in full in LICENSE.md; by downloading them you agree to its terms, including its restrictions on military, weapons-related and other prohibited uses.

Files

The .onnx.data sidecar is mandatory — torch's exporter externalizes weights regardless of the 2 GB protobuf limit, so the .onnx alone contains no weights. Keep the two files in the same directory; the ONNX parser resolves the sidecar next to the graph.

Model I/O

Input input [1, 3, 336, 336], ImageNet-normalized CHW (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225]), 336 = 21 × 16.

Verified at export: cos(pooler_output, last_hidden_state[:,0]) = 1.000000, so the descriptor is the CLS token. Note the 4 register tokens between CLS and the patch grid — any hand-rolled pooling must skip indices 1–4 or it averages in artifact tokens.

Exported with opset 18 from torch 2.11 / transformers 4.57.6 by crates/vrt-dinov3/scripts/export_dinov3.py.

⚠️ Build BF16, not FP16

FP16 produces all-NaN for this model. Measured on Jetson Orin (SM87, TensorRT 10.3.0.30), descriptor cosine against the PyTorch reference:

Cause: exactly one tensor leaves fp16 range — the attention logits Q·Kᵀ, max |value| 2.1e6 (2.5e6 on a natural image) against fp16's 65,504 ceiling. They overflow to inf in the first block, softmax(inf − inf) is NaN, and it reaches every output. Every other tensor stays under 1800. These are already-scaled logits — DINOv3 genuinely produces massive attention values (the "attention sink" its register tokens absorb). It is a dynamic-range problem, so bf16 fixes it by keeping fp32's exponent.

Do not try to pin fp16 with --layerPrecisions. TensorRT fuses the attention block into a myelin kernel, so ONNX node names are absent from the engine; both wildcard and explicit pins match nothing and are silently ignored — the build succeeds, runs at identical speed, and still emits NaN.

trtexec --onnx=dinov3-vits16-336.onnx --saveEngine=dinov3.engine --bf16 \
        --memPoolSize=workspace:2048

Usage

let dino = DinoV3::from_hub(stream.clone())?;   // or from_engine_file / from_onnx
let mut r = dino.alloc_result()?;
dino.submit(&img, &mut r)?;                      // enqueue, no sync
stream.synchronize()?;
let d = r.descriptor_host()?;                    // L2-normed [384], cosine-ready

Measured separation (real images): two views of one place score ~0.96; unrelated scenes −0.001 … 0.11. Live on a 1280×720 RTSP camera: 15 fps, source-gated, with the GPU using 6.8 ms of the 66 ms frame interval.

Citation

@article{simeoni2025dinov3,
  title={DINOv3},
  author={Sim{\'e}oni, Oriane and others},
  journal={arXiv preprint arXiv:2508.10104},
  year={2025}
}