F-Clip — Fully Convolutional Line Parsing

This repository contains the official PyTorch implementation of the paper: *Xili Dai, Xiaojun Yuan, Haigang Gong, Yi Ma. "Fully Convolutional Line Parsing." *.

Our method (F-Clip) is a simple and effective neural network for detecting the line from a given image and video. It outperforms the previous state-of-the-art wireframe and line detectors by a large margin on both accuracy and speed. We hope that this repository serves as a new reproducible baseline for future researches in this area.

Main results


The accuracy and speed trade-off among most recent wireframe detection methods on ShanghaiTech dataset

Qualitative Measures


More random sampled results can be found in the paper.

Quantitative Measures

The following table reports the performance metrics of several wireframes and line detectors on the ShanghaiTech dataset. A detail leader board of the ShanghaiTech and YorkUrban datasets can be found here.


Reproducing Results


For the ease of reproducibility, you are suggested to install miniconda (or anaconda if you prefer) before following executing the following commands.

git clone https://github.com/Delay-Xili/F-Clip
cd F-Clip
conda create -y -n fclip
source activate fclip
# Replace cudatoolkit=10.1 with your CUDA version: https://pytorch.org/
conda install -y pytorch cudatoolkit=10.1 -c pytorch
conda install -y pyyaml docopt matplotlib scikit-image opencv
mkdir data logs post

Testing Pre-trained Models

You can download our reference 6 pre-trained models HG1_D2, HG1_D3, HG1, HG2, HG2_LB, and HR from Google Drive. Those models were trained with their corresponding settings config/fclip_xxx.yaml.
To generate wireframes on the validation dataset with the pretrained model, execute

python test.py -d 0 -i <directory-to-storage-results> config/fclip_xxx.yaml <path-to-xxx-ckpt-file> shanghaiTech/york <path-to-validation-set>

Detect Wireframes for Your Own Images or Videos

To test F-Clip on your own images or videos, you need to download the pre-trained models and execute

CUDA_VISIBLE_DEVICES=0 python demo.py <path-to-image-or-video> --model HR --output_dir logs/demo_result --ckpt <path-to-pretrained-pth> --display True

Here, --output_dir is specifying the directory where the results will store, and you can specify --display to see the results on time.

Downloading the Processed Dataset

You can download the processed dataset wireframe.zip and york.zip manually from Google Drive (link1, link2).

Processing the Dataset

Optionally, you can pre-process (e.g., generate heat maps, do data augmentation) the dataset from scratch rather than downloading the processed one.

dataset/wireframe.py data/wireframe_raw data/wireframe
dataset/wireframe_line.py data/wireframe_raw data/wireframe


To evaluate the sAP (recommended) of all your checkpoints under logs/, execute

python eval-sAP.py logs/*/npz/*

MATLAB is required for APH evaluation and matlab should be under your $PATH. The parallel computing toolbox is highly suggested due to the usage of parfor. After post processing, execute

python eval-APH.py pth/to/input/npz pth/to/output/dir

Due to the usage of pixel-wise matching, the evaluation of APH may take up to an hour depending on your CPUs. See the source code of eval-sAP.py, eval-APH.py, and FClip/postprocess.py for more details on evaluation.


To train the neural network on GPU 0 (specified by -d 0) with the different 6 parameters, execute

python train.py -d 0 -i HG1_D2 config/fclip_HG1_D2.yaml
python train.py -d 0 -i HG1_D3 config/fclip_HG1_D3.yaml
python train.py -d 0 -i HG1 config/fclip_HG1.yaml
python train.py -d 0 -i HG2 config/fclip_HG2.yaml
python train.py -d 0 -i HG2_LB config/fclip_HG2_LB.yaml
python train.py -d 0 -i HR config/fclip_HR.yaml


We thank Yichao Zhou and Haozhi Qi of Berkeley for their help during all processes include ideas and experiments. This project would not have been possible without their help.


If you find F-Clip useful in your research, please consider citing:

      title={Fully Convolutional Line Parsing}, 
      author={Xili Dai and Xiaojun Yuan and Haigang Gong and Yi Ma},