Validate the MobileSAM segmentation result

Reconstruct the target mask

Run the visualization tool from the ExecuTorch repository root. The tool decodes the mask from the Fixed Virtual Platform (FVP) log and compares it with the host quantized mask:

    

        
        
python examples/arm/mobilesam_prompt_segmentation_example_ethos_u/runtime/visualize_fvp_output.py \
  --fvp-log=arm_test/mobilesam_manual/fvp.log \
  --input-image=examples/models/dinov2/dog.jpg \
  --metadata=arm_test/mobilesam_manual/export/mobilesam_point_ethos_u85_448.json \
  --reference-mask=arm_test/mobilesam_manual/export/debug/dog/quantized_mask.png \
  --minimum-iou=0.9 \
  --output-dir=arm_test/mobilesam_manual/fvp_visual

    

The command reports the number of foreground pixels and the agreement between the FVP and host masks. It exits with an error if the mask is empty, full, or below 0.9 intersection over union (IoU).

For the tested revision, the output is similar to:

    

        
        FVP mask: 3234 foreground pixels, artifacts saved to arm_test/mobilesam_manual/fvp_visual
FVP/reference IoU=0.9809 agreement=0.9951

        
    

Inspect the metrics

Display the target comparison metrics:

    

        
        
python -m json.tool arm_test/mobilesam_manual/fvp_visual/metrics.json

    

Confirm that fvp_reference_iou is at least 0.9. The fvp_reference_pixel_agreement value reports the fraction of matching pixels.

The exporter already checked the floating-point and quantized host masks. Display that result if you want to inspect the earlier stage:

    

        
        
python -m json.tool \
  arm_test/mobilesam_manual/export/mobilesam_point_ethos_u85_448_metrics.json

    

The fp32_quantized_mean_iou value must also be at least 0.9.

Confirm Ethos-U delegation

Display the delegation report written during export:

    

        
        
sed -n '1,120p' \
  arm_test/mobilesam_manual/export/mobilesam_point_ethos_u85_448_delegation.txt

    

The tested revision report is similar to:

    

        
        Total delegated subgraphs: 1
Number of delegated nodes: 5096
Number of non-delegated nodes: 3

        
    

Confirm that the graph is represented by one Ethos-U delegate. The three non-delegated graph nodes are expected at the delegate boundary. This count doesn’t mean that three model operators execute on the CPU.

Inspect the visual result

Open arm_test/mobilesam_manual/fvp_visual/fvp_comparison.png in an image viewer. The image contains three panels:

• The resized input image with the positive point prompt
• The host quantized segmentation overlay
• The FVP segmentation overlay

Compare the object boundaries in the two mask overlays. The FVP mask should select the dog at the positive point prompt and closely match the host quantized result.

What you’ve accomplished

You’ve completed the MobileSAM deployment flow from PyTorch export to bare-metal execution on Ethos-U85. You also confirmed that the target produces a non-degenerate mask that agrees with the host quantized reference.

You can now use this workflow to export, deploy, and validate quantized MobileSAM prompt segmentation models on Arm Ethos-U85 for your own use cases.