VMD-Net — Video Mirror Detection Network

Pre-trained weights for VMD-Net, introduced in:

Learning to Detect Mirrors from Videos via Dual Correspondences
Jiaying Lin*, Xin Tan*, Rynson W. H. Lau
CVPR 2023
Paper · Project Page · Dataset (VMD-D)

Model Summary

VMD-Net detects mirrors in video sequences by exploiting dual correspondences — both intra-frame (spatial) and inter-frame (temporal) — via a Relation Attention module built on a DeepLabV3 encoder backbone. This design lets the model handle frames where intra-frame mirror cues are weak or absent, producing accurate and temporally consistent segmentation masks.

File	Description
best.pth	Best checkpoint (714 MB), saved as {'model': state_dict, ...}
results/results.zip	VMD-Net predictions on the VMD-D test set
results/baseline_results.zip	Baseline method predictions for comparison

Loading the Weights

import torch
from networks.VMD_network import VMD_Network   # from the code release

model = VMD_Network()
checkpoint = torch.load("best.pth", map_location="cpu")
model.load_state_dict(checkpoint["model"])
model.eval()

Download the checkpoint:

huggingface-cli download garrying/VMD-Net best.pth --local-dir ./weights

Training Dataset

This model was trained and evaluated on VMD-D, the first large-scale video mirror detection dataset:

• 14,987 frames from 269 videos with manually annotated binary masks
• Available at garrying/VMD-D

Citation

@InProceedings{Lin_2023_CVPR,
  author    = {Lin, Jiaying and Tan, Xin and Lau, Rynson W.H.},
  title     = {Learning To Detect Mirrors From Videos via Dual Correspondences},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2023},
  pages     = {9109-9118}
}

License

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