# Occluder Segmentation Models

This project supports loading multiple U²-Net variants and ensembling their
outputs. Place your ONNX checkpoints in this directory and restart the server.

## Recommended setup

1. Download the original full-size U²-Net model (`u2net.onnx`). The official
   release is available on GitHub:

   ```bash
   curl -L -o models/u2net.onnx \
     -H "Accept: application/octet-stream" \
     https://github.com/xuebinqin/U-2-Net/releases/download/v1.0/u2net.onnx
   ```

   > If GitHub rate-limits anonymous downloads, supply a token:
   >
   > ```bash
   > GITHUB_TOKEN=... \
   > curl -L -o models/u2net.onnx \
   >   -H "Authorization: token ${GITHUB_TOKEN}" \
   >   -H "Accept: application/octet-stream" \
   >   https://github.com/xuebinqin/U-2-Net/releases/download/v1.0/u2net.onnx
   > ```

2. Keep the lightweight `u2netp.onnx` alongside it. When
   `U2NET_USE_ALL_MODELS=1` (enabled in `ecosystem.config.js`), the backend
   loads every listed checkpoint and merges their masks.

3. To evaluate fine-tuned models, drop the ONNX file here and append its name to
   `U2NET_MODEL_PRIORITY` (comma-separated). For example:

   ```bash
   pm2 set nextjs:U2NET_MODEL_PRIORITY "custom_finetune.onnx,u2net.onnx,u2netp.onnx"
   ```

After updating the model set, rebuild and restart:

```bash
npm run build
npm run deploy
```

Use the visualizer’s “Source” dropdown to compare Processed / Server / Client
masks and quantify the impact of new checkpoints quickly.