LEDNet

Why we included this project

If you build perception systems that need to run on embedded or mobile hardware, this compact reference is worth studying. LEDNet is a semantic-segmentation network squeezed into under a million parameters, using channel split-and-shuffle residual blocks and an attention pyramid decoder to keep both compute and accuracy in check. The authors report over 71 FPS on a GTX 1080Ti and a favorable speed-accuracy trade-off on Cityscapes, and the repo ships the PyTorch training and evaluation code behind those numbers. That makes it a handy starting point for reproducing the paper's results, comparing lightweight backbones for urban scene parsing, or adapting a proven design pattern when you need dense pixel classification that stays cheap. Just note the code targets an older PyTorch stack, so you will likely need to port it before running on a current environment.