---
license: cc-by-nc-4.0
tags:
  - mirror-detection
  - image-segmentation
  - computer-vision
  - pytorch
---

# PMDNet — Progressive Mirror Detection

Pretrained weights for **PMDNet**, the model introduced in the CVPR 2020 paper [*Progressive Mirror Detection*](https://jiaying.link/cvpr2020-pgd/).

## Model Description

PMDNet progressively detects mirror surfaces by leveraging multi-scale contrast cues and relational context.

**Architecture overview:**

- **Backbone** — ResNeXt-101 (32×4d), producing feature maps at four scales.
- **Contrast Module** (`Contrast_Module_Deep`) — at each scale, dilated convolutions capture local–context differences, then four stacked `Contrast_Block_Deep` units compute pairwise local–context subtractions at two dilation rates. Outputs are aggregated with CBAM (channel + spatial attention).
- **Relation Attention** (`Relation_Attention` / `RAttention`) — criss-cross attention over rows, columns, and both diagonals, enabling long-range relational reasoning without a full self-attention map.
- **Decoder** — four transposed-convolution upsampling stages with CBAM refinement produce intermediate saliency predictions (`f4 → f1`), each gated by the previous scale's prediction for progressive focus.
- **Edge Branch** — extracts edge features from `layer1` fused with high-level `cbam_4` context, producing an explicit edge map.
- **Refinement** — a single 1×1 conv fuses the original image, all four scale predictions, and the edge map into the final mirror mask.

**Input:** RGB image, resized to 416×416.  
**Output (eval):** `(f4, f3, f2, f1, edge, final)` — sigmoid-activated predictions at input resolution.  
Optional CRF post-processing is applied to the final prediction.

## Weights

| File | Size | Description |
|------|------|-------------|
| `pmd.pth` | ~414 MB | Full model weights (ResNeXt-101 backbone + decoder) |

## Usage

```python
import torch
from torchvision import transforms
from PIL import Image
from model.pmd import PMD  # from the official code release

model = PMD()
model.load_state_dict(torch.load("pmd.pth", map_location="cpu"))
model.eval()

transform = transforms.Compose([
    transforms.Resize((416, 416)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

img = Image.open("your_image.jpg").convert("RGB")
x = transform(img).unsqueeze(0)

with torch.no_grad():
    f4, f3, f2, f1, edge, final = model(x)
# `final` is the mirror mask prediction (values in [0, 1])
```

Full inference script with CRF post-processing: see [`code_minimal/infer.py`](https://jiaying.link/cvpr2020-pgd/).

## Dataset

Trained on the [PMD dataset](https://huggingface.co/datasets/garrying/PMD) (5,095 training images with mirror masks and edge maps).

## Performance

| Method    | F_β   | MAE   |
|-----------|-------|-------|
| EGNet     | 0.672 | 0.087 |
| MirrorNet | 0.748 | 0.061 |
| **PMDNet (ours)** | **0.790** | **0.032** |

Evaluated on the PMD test split (571 images).

## License

CC BY-NC 4.0 — non-commercial use only.

## Citation

```bibtex
@INPROCEEDINGS{PMD:2020,
   Author    = {Jiaying Lin and Guodong Wang and Rynson W.H. Lau},
   Title     = {Progressive Mirror Detection},
   Booktitle = {Proc. CVPR},
   Year      = {2020}
}
```

## Contact

csjylin@gmail.com