---
license: mit
tags:
- line-segment-detection
- deeplsd
- image-features
---

# DeepLSD — weight mirror

Unmodified mirror of the two checkpoints published for
[DeepLSD](https://github.com/cvg/DeepLSD) (CVPR 2023, Pautrat et al.), a hybrid line segment
detector: a network predicts a line distance and angle field, and a modified LSD reads the
segments off those predicted fields instead of off the image gradient.

Upstream ships no Hugging Face repo and serves these files from a university web server. This
repo exists so they can be fetched with `huggingface_hub` and loaded offline afterwards.

**Nothing here is retrained, converted or quantised** — the files are byte-identical to the
upstream artefacts listed below. (They are `torch.save` archives despite the `.tar` extension,
which is upstream's naming.)

## Contents

| File | Bytes | SHA-256 |
|------|------:|---------|
| `deeplsd_md.tar` | 102,898,193 | `56e9bae263977caa289ae49802987a67d4f00f376ae7e542176c8ae0a6cc2083` |
| `deeplsd_wireframe.tar` | 102,898,321 | `6d0c1bc78f79eb9e564404e4fef797d12ab18f30ba596208cca16116c2429406` |

## Which one to use

| File | Trained on / suited to |
|------|------------------------|
| `deeplsd_md.tar` | Trained on MegaDepth. The generic checkpoint: outdoor scenes and harder images (night, fog, blur). Upstream's own demo default. |
| `deeplsd_wireframe.tar` | Trained on the Wireframe dataset. Clean man-made scenes -- documents, floorplans, CAD exports, indoor rooms. |

## Provenance

| File | Upstream source |
|------|-----------------|
| `deeplsd_md.tar` | `https://cvg-data.inf.ethz.ch/DeepLSD/deeplsd_md.tar` |
| `deeplsd_wireframe.tar` | `https://cvg-data.inf.ethz.ch/DeepLSD/deeplsd_wireframe.tar` |

Each file was verified before upload by loading it into DeepLSD's own network and requiring an
exact fit — zero missing and zero unexpected `state_dict` keys.

## Usage

```python
import torch
from huggingface_hub import hf_hub_download

# DeepLSD ships no pip package; see the upstream repo for the model definition.
from deeplsd.models.deeplsd_inference import DeepLSD

path = hf_hub_download("xboix/deeplsd", "deeplsd_md.tar")
net = DeepLSD({"detect_lines": True})
net.load_state_dict(torch.load(path, map_location="cpu")["model"])
net = net.eval()
```

## Citation

```bibtex
@InProceedings{pautrat_suarez_2023_deeplsd,
  title={DeepLSD: Line Segment Detection and Refinement with Deep Image Gradients},
  author={Pautrat, R{\'e}mi and Barath, Daniel and Larsson, Viktor and Oswald, Martin R. and Pollefeys, Marc},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
  year={2023},
}
```