u2net-ort / README.md
ipsilondev's picture
Initial upload
6fc8e5a verified
|
Raw History Blame Contribute Delete
1.83 kB
metadata
license: apache-2.0
base_model: xuebinqin/U-2-Net
library_name: onnxruntime
pipeline_tag: image-segmentation
tags:
  - background-removal
  - salient-object-detection
  - onnx
  - ort-format

u2net-ort

U2-Net (full u2net, 320x320) in ORT format, dynamically quantised. Upstream: https://github.com/xuebinqin/U-2-Net (Apache-2.0). This is a MOVE of the file already shipped by the EraseBGAI app (private repo shadowofsoul/erasebgai, path EraseBGAI/erasebgaimodels/src/main/assets/models/u2net.quantz.ort). The bytes are identical to that location.

File Bytes SHA-256
u2net.quantz.ort 44741000 fce9e954a3a94921c860557353235931bbdacdb89930c223449af3c549315a3f

Tensors (checked with onnxruntime 1.24.1, CPU)

Name Shape dtype
input input.1 [1, 3, 320, 320] (NCHW) float32
outputs 1959 ... 1965 (7 side outputs d0..d6) [1, 1, 320, 320] each float32

Use output 0 (1959, the fused map d0). Values are already in 0..1 (sigmoid applied inside the graph).

Pre- and post-processing (as in EraseBGAI utils.java: processImageToInputU2net / ouputTensorToImageU2net)

  • Resize to 320x320 (bilinear), x / 255, normalise with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225], NCHW.
  • Output: take output 0, min-max normalise to 0..1 over the whole map ((v - min) / (max - min)), resize back to the source size. Result is the alpha matte.

Provenance

The original conversion command is not recorded in the EraseBGAI repository. The file name (quantz) and the ORT flatbuffer header (ORTM) indicate: public U2-Net PyTorch weights -> ONNX export -> dynamic quantisation -> convert_onnx_models_to_ort. The graph signature above matches the standard U2-Net ONNX export (input.1, seven outputs).