license: apache-2.0
library_name: onnx
tags:
- onnx
- object-detection
- image-feature-extraction
- webgpu
inframe browser models
The three files the inframe web page downloads to cut a video into the clips where a chosen target is in frame, entirely in the visitor's browser. The video never leaves the browser; only these files are fetched, once per browser.
| file | role | bytes | sha256 | license | origin |
|---|---|---|---|---|---|
dinov2_s_v1.onnx |
embedder | 88,405,648 | 736056bcbb7f117b7d64a56f1588f0986d1667f1d91b6e851f854e207a05c22f |
Apache-2.0 | https://huggingface.co/facebook/dinov2-small (pooler_output, opset 17) |
ort-wasm-simd-threaded.asyncify.wasm |
runtime | 26,781,914 | 39f9f0894d478800487ed9f7dbe92618498db320cf55c8e3d89adff8dce658da |
MIT | https://www.npmjs.com/package/onnxruntime-web (1.30.0, WebGPU build, ort.webgpu.bundle) |
rfdetr_small_v1.onnx |
detector | 123,510,543 | 246ed1ed4eae0e53fac52352f02866308b7a4291f5ab002320ab9f7744e7e07b |
Apache-2.0 | https://github.com/roboflow/rf-detr (RF-DETR small, COCO checkpoint, rfdetr 1.5.2 export) |
manifest.json holds the same digests. The page carries its own copy of the manifest and checks every file against it before using it, so a
changed file is dropped.
Licenses
rfdetr_small_v1.onnx and dinov2_s_v1.onnx are Apache-2.0 (LICENSE-APACHE); ort-wasm-simd-threaded.asyncify.wasm is MIT (LICENSE-MIT).
NOTICE says where each file comes from and what was changed (a format conversion only, no retraining). The card header says apache-2.0 because
the two weight files are; the runtime file keeps its own MIT license. The wasm runtime is a statically linked binary that also contains
abseil-cpp and onnx (Apache-2.0), protobuf (BSD-3-Clause), Eigen (MPL-2.0) and a few smaller libraries; their notices and license texts are in
THIRD_PARTY_NOTICES.
How they were made
tools/export_onnx.py in the inframe repository (RF-DETR small with rfdetr 1.5.2, DINOv2-S from facebook/dinov2-small at opset 17, fp32, and a copy of
the onnxruntime-web 1.30.0 asyncify wasm):
INFRAME_MODELS=... python tools/export_onnx.py
Use
Fetch a file from a fixed commit (/resolve/<40-hex commit>/<file>), never from main, and compare its sha256 with manifest.json.