---
license: apache-2.0
tags:
  - segmentation
  - prompted-segmentation
  - clipseg
  - drywall
---
# drywall-clipseg

CLIPSeg fine-tune (HF `CIDAS/clipseg-rd64-refined` backbone, decoder + FiLM
unfrozen) for prompted binary segmentation on drywall imagery. Single
checkpoint covers two classes selected by text prompt:

- `"segment crack"` → wall-crack mask
- `"segment taping area"` → drywall taping-seam mask

## Test metrics (focal_dice loss, threshold 0.6)

| Task | Dice | mIoU | Precision | Recall |
| --- | --- | --- | --- | --- |
| Crack  | 0.672 | 0.531 | — | — |
| Taping | 0.727 | 0.587 | — | — |

## Load + predict

```python
from huggingface_hub import hf_hub_download
import torch
from src.models.clipseg_wrapper import CLIPSegFT

ckpt_path = hf_hub_download(repo_id="ravindrakapse/drywall-clipseg", filename="best.pt")
model = CLIPSegFT(pretrained="CIDAS/clipseg-rd64-refined").cuda()
state = torch.load(ckpt_path, map_location="cuda")
model.load_state_dict(state["model"])
model.eval()
```

See [`load_models.py`](https://github.com/Ravindrakapse/prompt_segmentation/blob/main/load_models.py) for the full inference pipeline (letterbox + TTA + un-letterbox).