After converting the SAM2 model to ONNX , the inference results are significantly worse than the original model.

[lhz123]

[cile98]

@lhz123 can you provide any comparison examples?

[vietanhdev]

Hi! :D The inference flow is not identical to the original one from SAM2 official repo, including the dimension of the input images. Therefore, the results cannot be comparable.

[marwand]

@vietanhdev I've noticed all of the converted SAM2 models output a mask in a 256x256 resolution. Is this configurable? Ideally I want it to be the same as the input resolution (1024x1024).

The reason 256 isn't good enough is that after upscaling to 1024, the edges are very rough and don't overlay perfectly with the source image. I've applied some basic post processing, but the result isn't very accurate, especially for small object/surfaces.

Does the original SAM model output masks in 256 res? What are the limitations that make the onnx version different from the pytorch one?

[cile98]

@vietanhdev I've noticed all of the converted SAM2 models output a mask in a 256x256 resolution. Is this configurable? Ideally I want it to be the same as the input resolution (1024x1024).

The reason 256 isn't good enough is that after upscaling to 1024, the edges are very rough and don't overlay perfectly with the source image. I've applied some basic post processing, but the result isn't very accurate, especially for small object/surfaces.

Does the original SAM model output masks in 256 res? What are the limitations that make the onnx version different from the pytorch one?

Pretty sure SAM1 also originally outputs them in 256x256 res and then upscales them

[ibaiGorordo]

@vietanhdev I recommend adding masks = F.interpolate(masks, (img_size[0], img_size[1]), mode="bilinear", align_corners=False) to the decoder to get smoother results than doing the upscale with opencv

here is the updated colab notebook for export: https://colab.research.google.com/drive/1tqdYbjmFq4PK3Di7sLONd0RkKS0hBgId?usp=sharing

[vietanhdev]

Hi @ibaiGorordo Thank you for your great code! Could you help with a PR to this repo?

[marwand]

@vietanhdev I recommend adding masks = F.interpolate(masks, (img_size[0], img_size[1]), mode="bilinear", align_corners=False) to the decoder to get smoother results than doing the upscale with opencv

here is the updated colab notebook for export: https://colab.research.google.com/drive/1tqdYbjmFq4PK3Di7sLONd0RkKS0hBgId?usp=sharing

This is great, thank you!