anakhiu/fastsam-onnx

FastSAM-s ONNX (fp32)

ONNX export of FastSAM-s (YOLOv8s-seg, ~11.8M params) for click-to-select segmentation on CPU. Used by MicrostockEngine to mask a watermark from a single click before inpainting.

Provenance

• Source weights: ultralytics FastSAM-s.pt, assets release v8.4.0.
• Export: ultralytics 8.4.107, imgsz=1024, opset=17, simplified with onnxslim.
• No quantization — full fp32.

I/O contract

• input images: [1, 3, 1024, 1024] float32, values 0..1, letterboxed top-left with pad value 114/255.
• output0: [1, 37, 21504] — per anchor: 4 box (xywh, in 1024-space) + 1 class score (sigmoid already applied; FastSAM is class-agnostic) + 32 mask coefficients.
• output1: [1, 32, 256, 256] — 32 mask prototypes at a quarter of input resolution. An instance mask is sigmoid(sum(coeff_i * proto_i)), cropped to its own box.

Licence

Apache-2.0 (CASIA-IVA-Lab/FastSAM). The weights were trained with Ultralytics code (AGPL-3.0); noted here so the provenance is not lost.