LINEA: Fast and accurate line detection using scalable transformers

📄 This is the official implementation of the paper:
LINEA: Fast and accurate line detection using scalable transformers

Sebastian Janampa and Marios Pattichis

The University of New Mexico
Department of Electrical and Computer Engineering

LINEA is a powerful real-time line detector that introduces Line Attention mechanism, achieving outstanding performance without being pretrained on COCO or Object365 datasets.

Attention Mechanishm

We compare line attention with traditional attention and deformable attention. We highlight two advantages of our proposed mechanism:

• Line attention is a sparse mechanism like deformable attention. This significantly reduces memory complexity.
• Line attention pays attention to the line endpoints like traditional attention but also attends locations between the endpoints.

🚀 Updates

• [2025.02.19] Release LINEA series.
• [2025.02.20] Release LINEA weights.
• [2025.02.20] Release Google Colab Notebook.
• [2025.02.20] Release HuggingFace 🤗 Space.

📝 TODO

• Upload paper (currently under review)
• Upload requirements
• Upload LINEA weigths.
• Create HuggingFace 🤗 space.
• Create Collab demo.

Model Zoo

Wireframe

Model	Dataset	AP5	AP10	AP15	#Params	Latency	GFLOPs	config	checkpoint
LINEA‑N	Wireframe	58.7	65.0	67.9	3.9M	2.54ms	12.1	py	65.0
LINEA‑S	Wireframe	58.4	64.7	67.6	8.6M	3.08ms	31.7	py	64.7
LINEA‑M	Wireframe	59.5	66.3	69.1	13.5M	3.87ms	45.6	py	66.3
LINEA‑L	Wireframe	61.0	67.9	70.8	25.2M	5.78ms	83.8	py	67.9

YorkUrban

Model	Dataset	AP5	AP10	AP15	#Params	Latency	GFLOPs	config	checkpoint
LINEA‑N	YorkUrban	27.3	30.5	32.5	3.9M	2.54ms	12.1	py	65.0
LINEA‑S	YorkUrban	28.9	32.6	34.8	8.6M	3.08ms	31.7	py	64.7
LINEA‑M	YorkUrban	30.3	34.5	36.7	13.5M	3.87ms	45.6	py	66.3
LINEA‑L	YorkUrban	30.9	34.9	37.3	25.2M	5.78ms	83.8	py	67.9

Notes:

• Latency is evaluated on a single NVIDIA RTX A5500 GPU with $batch\_size = 1$, $fp16$, and $TensorRT==10.5.0$.

Quick start

Setup

conda create -n linea python=3.11.9
conda activate linea
pip install -r requirements.txt

Data Preparation

Wireframe Dataset

TODO

YorkUrban Dataset

TODO

Usage

Wireframe

1. Set Model

export model=l  # n s m l

2. Training

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=7777 --nproc_per_node=4 main.py -c configs/linea/linea_hgnetv2_${model}.py --coco_path data/wireframe_processed --amp 

3. Testing

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_${model}.py --coco_path data/wireframe_processed --amp  --eval --resume <checkpoit.pth>

4. Replicate results (optional)

# First, download the official weights
wget https://github.com/SebastianJanampa/storage/releases/download/LINEA/linea_hgnetv2_${model}.pth

# Second, run test
CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_${model}.py --coco_path data/wireframe_processed --amp  --eval --resume linea_hgnetv2_${model}.pth

YorkUrban

1. Set Model

export model=l  # n s m l

2. Testing

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_${model}.py --coco_path data/york_processed --amp  --eval --resume linea_hgnetv2_${model}.pth

3. Replicate results (optional)

# First, download the official weights
wget https://github.com/SebastianJanampa/storage/releases/download/LINEA/linea_hgnetv2_${model}.pth

# Second, run test
CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_${model}.py --coco_path data/york_processed --amp  --eval --resume linea_hgnetv2_${model}.pth

Customizing Batch Size

For example, if you want to train with a total batch size of 16 when training LINEA-L on Wireframe:

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=7777 --nproc_per_node=4 main.py -c configs/linea/linea_hgnetv2_l.py --coco_path data/wireframe_processed --amp --options batch_size_train=16

Customizing Input Size

If you'd like to train LINEA-L on Wireframe with an input size of 320x320 (we only support square shapes):

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_l.py --coco_path data/wireframe_processed --amp --options eval_spatial_size=320,320

or

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_l.py --coco_path data/wireframe_processed --amp --options eval_spatial_size=320

Multiple Costumizations

If you'd like to train LINEA-L on Wireframe with an input size of 480x480 and a total batch size of 4:

CUDA_VISIBLE_DEVICES=0 torchrun --master_port=7777 --nproc_per_node=1 main.py -c configs/linea/linea_hgnetv2_l.py --coco_path data/wireframe_processed --amp --options eval_spatial_size=320 batch_size_train=4

Tools

Deployment

1. Setup

pip install onnx onnxsim
export model=l  # n s m l

2. Export onnx

python tools/deployment/export_onnx.py --check -c configs/linea/linea_hgnetv2_${model}_coco.yml -r linea_hgnetv2_${model}.pth

3. Export tensorrt For a specific file

trtexec --onnx="model.onnx" --saveEngine="model.engine" --fp16

or, for all files inside a folder

python export_tensorrt.py

Inference (Visualization)

1. Setup

pip install -r tools/inference/requirements.txt
export model=l  # n s m l

2. Inference (onnxruntime / tensorrt / torch)

Inference on images and videos is supported.

For a single file

python tools/inference/onnx_inf.py --onnx model.onnx --input example/example1.jpg  
python tools/inference/trt_inf.py --trt model.engine --input example/example1.jpg
python tools/inference/torch_inf.py -c configs/linea/linea_hgnetv2_${model}.yml -r <checkpoint.pth> --input example/example1.jpg --device cuda:0

For a folder

python tools/inference/onnx_inf.py --onnx model.onnx --input example  
python tools/inference/trt_inf.py --trt model.engine --input example
python tools/inference/torch_inf.py -c configs/linea/linea_hgnetv2_${model}.yml -r linea_hgnetv2_${model}t.pth --input example --device cuda:0

Benchmark

1. Setup

pip install -r tools/benchmark/requirements.txt
export model=l  # n s m l

2. Model FLOPs, MACs, and Params

python tools/benchmark/get_info.py --config configs/linea/linea_hgnetv2_${model}.py 

3. TensorRT Latency

python tools/benchmark/trt_benchmark.py --infer_dir ./data/wireframe_processed/val2017 --engine_dir trt_engines

4. Pytorch Latency

python tools/benchmark/torch_benchmark.py -c ./configs/linea/linea_hgnetv2_${model}.py --resume linea_hgnetv2_${model}.pth --infer_dir ./data/wireframe_processed/val2017

Visualization: Line attention

python tools/visualization/line_attention.py -c ./configs/linea/linea_hgnetv2_${model}.py --resume linea_hgnetv2_${model}.pth --data-path ./data/wireframe_processed -d cuda --num_images 10

Visualization: Feature maps from the backbone and encoder

python tools/visualization/backbone_encoder.py -c ./configs/linea/linea_hgnetv2_${model}.py --resume linea_hgnetv2_${model}.pth --data-path ./data/wireframe_processed -d cuda --num_images 10

Citation

If you use LINEA or its methods in your work, please cite the following BibTeX entries:

bibtex

TODO

Acknowledgement

Our work is built upon DT-LSD and D-FINE. Thanks to the inspirations from DT-LSD, D-FINE, RT-DETR, and LETR.

✨ Feel free to contribute and reach out if you have any questions! ✨