Understand the MobileSAM deployment workflow

What you’ll run

MobileSAM is a compact variant of the Segment Anything Model. It takes an image and point prompts, then predicts a binary mask for the object selected by those points.

You’ll run the ExecuTorch MobileSAM example on a Corstone-320 Fixed Virtual Platform (FVP). The system combines a Cortex-M85 processor with an Ethos-U85 neural processing unit (NPU). This lets you test the complete bare-metal deployment without a physical board.

Follow the model from PyTorch to the target

You’ll run an ExecuTorch MobileSAM example that performs the following sequence:

1. Downloads the pinned MobileSAM vit_t source and checkpoint.
2. Embeds the positive point (219, 193) into the exported model.
3. Calibrates post-training quantization with the example dog image.
4. Uses 8-bit activations and weights for most of the graph, with 16-bit activations and 8-bit weights for TinyViT attention.
5. Lowers the quantized graph to an Ethos-U85-256 delegate and writes an ExecuTorch .pte program.
6. Builds the standard Arm ExecuTorch runner, runs it on the Corstone-320 FVP, and compares its mask with the host quantized mask.

The default image tensor has shape [1, 3, 448, 448]. The model produces one mask-logit tensor with shape [1, 1, 112, 112].

Understand the fixed-prompt contract

The exported .pte accepts an image tensor as its only runtime input. The runner embeds the .pte and reads the preprocessed image from input.bin through semihosting, which lets the FVP access files on your host. It writes the output tensor to output-0.bin through the same mechanism.

You can supply a different image tensor with the same shape and preprocessing without exporting again or rebuilding the runner. The fixed point coordinates must still identify the intended object in the resized and padded image. Validation also needs the host quantized mask and visualization image for that same input. Changing the point coordinates requires a new export and a rebuild of the runner that embeds the program.

The example provides one tested configuration: MobileSAM vit_t, a 448x448 image, point (219, 193), and Ethos-U85-256. The MobileSAM preparation, export, and visualization scripts use fixed settings and paths, so you invoke them without configuration arguments.

The example uses multimask_output=False. The target returns low-resolution mask logits; the host thresholds them at zero and resizes the binary mask for visualization.

Know what the validation proves

The workflow performs two comparisons:

• Host validation compares the floating-point mask with the quantized mask before lowering.
• Target validation compares the mask produced by the FVP with the host quantized mask.

Both comparisons enforce a minimum intersection over union (IoU) of 0.9. Export stops if the host comparison falls below that threshold. The visualization step checks the output tensor size and stops if the target comparison falls below 0.9.

These checks validate agreement across the deployment stages for the example image. The same dog image is used for calibration and evaluation. For an application, use a representative calibration set and labeled test images to assess segmentation quality. The fast FVP mode used here validates correctness; use hardware measurements to assess real-device latency.

What you’ve learned and what’s next

You now know what you’ll deploy, why the point prompt is fixed, and how the host and target checks cover different stages of the pipeline.

Next, you’ll prepare ExecuTorch and the Arm development tools.