/LETR

(CVPR 2021 Oral) LETR: Line Segment Detection Using Transformers without Edges

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

LETR: Line Segment Detection Using Transformers without Edges

Introduction

This repository contains the official code and pretrained models for Line Segment Detection Using Transformers without Edges. Yifan Xu*, Weijian Xu*, David Cheung, and Zhuowen Tu. CVPR2021 (Oral)

In this paper, we present a joint end-to-end line segment detection algorithm using Transformers that is post-processing and heuristics-guided intermediate processing (edge/junction/region detection) free. Our method, named LinE segment TRansformers (LETR), takes advantages of having integrated tokenized queries, a self-attention mechanism, and encoding-decoding strategy within Transformers by skipping standard heuristic designs for the edge element detection and perceptual grouping processes. We equip Transformers with a multi-scale encoder/decoder strategy to perform fine-grained line segment detection under a direct endpoint distance loss. This loss term is particularly suitable for detecting geometric structures such as line segments that are not conveniently represented by the standard bounding box representations. The Transformers learn to gradually refine line segments through layers of self-attention.

Model Pipeline

Changelog

05/07/2021: Code for LETR Basic Usage Demo are released.

04/30/2021: Code and pre-trained checkpoint for LETR are released.

Results and Checkpoints

Name sAP10 sAP15 sF10 sF15 URL
Wireframe 65.6 68.0 66.1 67.4 LETR-R101
Wireframe 64.6 67.3 64.8 66.2 LETR-R50
YorkUrban 27.5 29.6 39.4 40.9 LETR-R101
YorkUrban 29.4 31.7 40.3 42.0 LETR-R50

Reproducing Results

Step1: Code Preparation

git clone https://github.com/mlpc-ucsd/LETR.git

Step2: Environment Installation

Add tsinghua mirror channel to anaconda and pip according to Anaconda 镜像使用帮助PyPI 镜像使用帮助

mkdir -p data evaluation/data exp

conda create -n letr python=3.8 anaconda         # There is a constraint between the Python and PyTorch versions
conda activate letr
# conda install -c pytorch pytorch torchvision
conda install cython opencv 
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Step3: Data Preparation

To reproduce our results, you need to process two datasets, ShanghaiTech and YorkUrban. Files located at ./helper/wireframe.py and ./helper/york.py are both modified based on the code from L-CNN, which process the raw data from download.

  • ShanghaiTech Train Data
    • To Download (modified based on from L-CNN)
      cd data
      bash ../helper/gdrive-download.sh 1BRkqyi5CKPQF6IYzj_dQxZFQl0OwbzOf wireframe_raw.tar.xz
      tar xf wireframe_raw.tar.xz
      rm wireframe_raw.tar.xz
      python ../helper/wireframe.py ./wireframe_raw ./wireframe_processed
      
  • YorkUrban Train Data
    • To Download
      cd data
      wget https://www.dropbox.com/sh/qgsh2audfi8aajd/AAAQrKM0wLe_LepwlC1rzFMxa/YorkUrbanDB.zip
      unzip YorkUrbanDB.zip 
      python ../helper/york.py ./YorkUrbanDB ./york_processed
      
  • Processed Evaluation Data
    bash ./helper/gdrive-download.sh 1T4_6Nb5r4yAXre3lf-zpmp3RbmyP1t9q ./evaluation/data/wireframe.tar.xz
    bash ./helper/gdrive-download.sh 1ijOXv0Xw1IaNDtp1uBJt5Xb3mMj99Iw2 ./evaluation/data/york.tar.xz
    tar -vxf ./evaluation/data/wireframe.tar.xz -C ./evaluation/data/.
    tar -vxf ./evaluation/data/york.tar.xz -C ./evaluation/data/.
    rm ./evaluation/data/wireframe.tar.xz
    rm ./evaluation/data/york.tar.xz

Step4: Train Script Examples

  1. Train a coarse-model (a.k.a. stage1 model).

    # Usage: bash script/*/*.sh [exp name]
    bash script/train/a0_train_stage1_res50.sh  res50_stage1 # LETR-R50  
    bash script/train/a1_train_stage1_res101.sh res101_stage1 # LETR-R101 
  2. Train a fine-model (a.k.a. stage2 model).

    # Usage: bash script/*/*.sh [exp name]
    bash script/train/a2_train_stage2_res50.sh  res50_stage2  # LETR-R50
    bash script/train/a3_train_stage2_res101.sh res101_stage2 # LETR-R101 
  3. Fine-tune the fine-model with focal loss (a.k.a. stage2_focal model).

    # Usage: bash script/*/*.sh [exp name]
    bash script/train/a4_train_stage2_focal_res50.sh   res50_stage2_focal # LETR-R50
    bash script/train/a5_train_stage2_focal_res101.sh  res101_stage2_focal # LETR-R101 

Step5: Evaluation

  1. Evaluate models.
    # Evaluate sAP^10, sAP^15, sF^10, sF^15 (both Wireframe and YorkUrban datasets).
    bash script/evaluation/eval_stage1.sh [exp name]
    bash script/evaluation/eval_stage2.sh [exp name]
    bash script/evaluation/eval_stage2_focal.sh [exp name]

Citation

If you use this code for your research, please cite our paper:

@InProceedings{Xu_2021_CVPR,
    author    = {Xu, Yifan and Xu, Weijian and Cheung, David and Tu, Zhuowen},
    title     = {Line Segment Detection Using Transformers Without Edges},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {4257-4266}
}

Acknowledgments

This code is based on the implementations of DETR: End-to-End Object Detection with Transformers.