S-4-G-4-R/clipseg-drywall-qa

CLIPSeg — Fine-tuned for Drywall QA

Fine-tuned version of CIDAS/clipseg-rd64-refined for text-conditioned binary segmentation of drywall defects.

Supported Prompts

Prompt	Target Region	Val mIoU	Val Dice
segment crack	Wall cracks	0.7352	0.8336
segment taping area	Joint / tape seam	0.4985	0.6256

Training Details

Setting	Value
Base model	CIDAS/clipseg-rd64-refined
Epochs	20
Batch size	4
Learning rate	1e-4 (AdamW)
Scheduler	CosineAnnealingLR
Loss	BCE 0.5 + Dice 0.5
Image size	352 × 352
Threshold	0.5
Seed	42
Hardware	Tesla T4 (Google Colab)
Train time	~65.3 min
Avg inference	13.0 ms / image

Datasets

• Dataset 1 — Taping area: Drywall-Join-Detect
• Dataset 2 — Cracks: Cracks

Quick Usage

import torch
from PIL import Image
from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation

processor = CLIPSegProcessor.from_pretrained("S-4-G-4-R/clipseg-drywall-qa")
model     = CLIPSegForImageSegmentation.from_pretrained("S-4-G-4-R/clipseg-drywall-qa")
model.eval()

image  = Image.open("your_image.jpg").convert("RGB")
prompt = "segment crack"   # or "segment taping area"

inputs = processor(
    text=prompt, images=image,
    return_tensors="pt", padding=True
)

with torch.no_grad():
    logits = model(**inputs).logits

mask = (torch.sigmoid(logits[0]) > 0.5).numpy()   # boolean H×W mask

Test Results (best checkpoint — epoch 15)

Metric	segment crack	segment taping area
mIoU	0.6900 (test) / 0.7352 (val)	0.4985 (val)
Dice	0.7957 (test) / 0.8336 (val)	0.6256 (val)