LPS-Net

Lightweight and Progressively-Scalable Networks for Semantic Segmentation

The original paper can be found here.

Comparisons of inference speed/accuracy tradeoff on Cityscapes validation set. Inference speed of LPS-Net (-S, -M, and -L) are measured on an NVIDIA GTX 1080Ti GPU with TensorRT.

Getting Started

Requiremenets

Package	Version
torch	1.9.0+cu111
torchvision	0.10.0+cu111
numpy	1.21.1
onnx	1.10.0
onnx-simplifier	0.3.6
Pillow	8.3.1
TensorRT	7.1.3.4

Evaluation in Command Line

To evaluate the LPS-Net-S on the Cityscapes validation set with "val_miou.py", first setup the Cityscapes dataset and update data path in "imagelist_val.txt"/"val_miou.py" and (if needed), then run:

python val_miou.py

The expected output is:

Total 500 images for validation.
LPS-Net-S on Cityscapes validation set: mean IoU=73.9%

Measure the Latency

To measure the latency of LPS-Net-S on your device with TensorRT in FP32 mode, run:

python latency.py

Please ensure the TensorRT has been correctly installed and configured.

Files in Repository

File	Content
val_miou.py	Evaluate the mean IoU performance of LPS-Net-S on Cityscapes validation set.
latency.py	Measure the latency of LPS-Net with TensorRT in FP32 mode.
lpsnet.py	Definitions and implementation of LPS-Net.
expand.py	Progressive expansion of LPS-Net.
LPS-Net-S.pth	Weights of LPS-Net-S. It is trained on the Cityscapes training set.
imagelist_val.txt	A list of image-label pairs on the Cityscapes validation set. It is utilized to evaluate mean IoU performace in "val_miou.py". Note that the label images should use official "trainId".

Citation

Cite as below if you find this repository is helpful:

@article{zhang2022lpsnet,
  title   = {Lightweight and Progressively-Scalable Networks for Semantic Segmentation},
  author  = {Zhang, Yiheng and Yao, Ting and Qiu, Zhaofan and Mei, Tao},
  journal = {ArXiv},
  year    = {2022},
  volume  = {abs/2207.13600}
}