---
library_name: pimm
tags:
  - particle-physics
  - lartpc
  - point-cloud
---

# PoLAr-MAE — semantic segmentation

[PoLAr-MAE](https://arxiv.org/abs/2502.02558) fine-tuned for 4-class LArTPC semantic segmentation (shower, track, Michel, delta); reproduces the paper mF1 ≈ 0.82.

- **pimm type:** `PoLArMAE-SemSeg` · 4 classes

> Faithful eval needs PoLAr-MAE preprocessing: `LogTransform(min_val=0.13)`, `energy_threshold=0.13`, `remove_low_energy_scatters=True`. Coordinate normalization is in-model.

## Loading

```python
import pimm
model = pimm.from_pretrained("hf://deeplearnphysics/polar-mae-semantic")
```

Architecture + hyper-parameters travel in `config.json`; weights are bitwise-identical to the original checkpoint.

## Provenance

Repackaged from the original [PoLAr-MAE](https://github.com/DeepLearnPhysics/PoLAr-MAE) release checkpoints into the [pimm](https://github.com/youngsm/particle-imaging-models) export format. Inherits the source repo license.