PicoSAM3 — IMX500 Checkpoints

Compiled deployment files for PicoSAM3, a lightweight segmentation model (~1.5 MB) distilled from SAM3 and designed to run entirely on the Sony IMX500 AI camera sensor.

For code, training, and full documentation see the GitHub repository: https://github.com/pbonazzi/picosam3

---

Files

File	Description
checkpoints/rpk/network.rpk	Compiled model for the IMX500 — load directly with picamera2
checkpoints/imx_out/packerOut.zip	Intermediate output from imxconv-pt (needed to re-run imx500-package)
checkpoints/imx_out/dnnParams.xml	IMX500 network parameters
checkpoints/PicoSAM3_student_epoch1.pt	PyTorch student checkpoint (distilled from SAM3)
checkpoints/PicoSAM3_epoch1.pt	PyTorch supervised baseline

---

Quick start on Raspberry Pi 5 + IMX500

git clone https://github.com/pbonazzi/picosam3
cd picosam3

# Download network.rpk from this HuggingFace repo into checkpoints/rpk/
huggingface-cli download pietrobonazzi/picosam3 \
    checkpoints/rpk/network.rpk \
    --local-dir .

uv sync
uv run python3 demo_imx500.py

Left-click and drag on the live window to set the segmentation ROI. Press r to reset, q to quit.