Holistically-Attracted Wireframe Parsing (CVPR 2020)

This is the official implementation of our CVPR paper.

[News] We experimentally provided an easy-to-install version for inference-only usage of HAWP, please checkout the inference branch for the details.

Highlights

• We propose a fast and parsimonious parsing method HAWP to accurately and robustly detect a vectorized wireframe in an input image with a single forward pass.
• The proposed HAWP is fully end-to-end.
• The proposed HAWP does not require the squeeze module.
• State-of-the-art performance on the Wireframe dataset and YorkUrban dataset.
• The proposed HAWP achieves 29.5 FPS on a GPU (Tesla V100) for 1-batch inference.

Quantitative Results

Wireframe Dataset

Method	Wireframe Dataset							FPS
	sAP5	sAP10	sAP15	msAP	mAPJ	APH	FH
LSD	/	/	/	/	/	55.2	62.5	49.6
AFM	18.5	24.4	27.5	23.5	23.3	69.2	77.2	13.5
DWP	3.7	5.1	5.9	4.9	40.9	67.8	72.2	2.24
L-CNN	58.9	62.9	64.9	62.2	59.3	80.3	76.9	15.6
						82.8†	81.3†
L-CNN (re-trained)	59.7	63.6	65.3	62.9	60.2	81.6	77.9	15.6
						83.7†	81.7†
HAWP (Ours)	62.5	66.5	68.2	65.7	60.2	84.5	80.3	29.5
						86.1†	83.1†

YorkUrban Dataset

Method	YorkUrban Dataset							FPS
	sAP5	sAP10	sAP15	msAP	mAPJ	APH	FH
LSD	/	/	/	/	/	50.9	60.1	49.6
AFM	7.3	9.4	11.1	9.3	12.4	48.2	63.3	13.5
DWP	1.5	2.1	2.6	2.1	13.4	51.0	61.6	2.24
L-CNN	24.3	26.4	27.5	26.1	30.4	58.5	61.8	15.6
						59.6†	65.3†
L-CNN (re-trained)	25.0	27.1	28.3	26.8	31.5	58.3	62.2	15.6
						59.3†	65.2†
HAWP (Ours)	26.1	28.5	29.7	28.1	31.6	60.6	64.8	29.5
						61.2†	66.3†

Installation (tested on Ubuntu-18.04, CUDA 10.0, GCC 7.4.0)

conda create -n hawp python=3.6
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

cd hawp
conda develop .

pip install -r requirement.txt
python setup.py build_ext --inplace

Quickstart with the pretrained model

python scripts/predict.py --img figures/example.png

Training & Testing

Data Preparation

• Download the Wireframe dataset and the YorkUrban dataset from their project pages.
• Download the JSON-format annotations (Google Drive).
• Place the images to "hawp/data/wireframe/images/" and "hawp/data/york/images/".
• Unzip the json-format annotations to "hawp/data/wireframe" and "hawp/data/york".

The structure of the data folder should be

data/
   wireframe/images/*.png
   wireframe/train.json
   wireframe/test.json
   ------------------------
   york/images/*.png
   york/test.json

Training

CUDA_VISIBLE_DEVICES=0, python scripts/train.py --config-file config-files/hawp.yaml

The best model is manually selected from the model files after 25 epochs.

Testing

CUDA_VISIBLE_DEVICES=0, python scripts/test.py --config-file config-files/hawp.yaml [optional] --display

The output results will be saved to OUTPUT_DIR/$dataset_name.json. The dataset_name should be wireframe_test or york_test.

Structural-AP Evaluation

• Run scripts/test.py to get the wireframe parsing results.
• Run scripts/eval_sap.py to get the sAP results

# example on the Wireframe dataset
python scripts/eval_sap.py --path outputs/hawp/wireframe_test.json --threshold 10

Citations

If you find our work useful in your research, please consider citing:

@inproceedings{HAWP,
title = "Holistically-Attracted Wireframe Parsing",
author = "Nan Xue and Tianfu Wu and Song Bai and Fu-Dong Wang and Gui-Song Xia and Liangpei Zhang and Philip H.S. Torr
",
booktitle = "IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
year = {2020},
}

Acknowledgment

We acknowledge the effort from the authors of the Wireframe dataset and the YorkUrban dataset. These datasets make accurate line segment detection and wireframe parsing possible.