File size: 3,330 Bytes
5fffb08 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 | ---
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
|