Abstract
In this project, we present ShelfNet, a lightweight convolutional neuralnetwork for accurate real-time semantic segmentation. Different from thestandard encoder-decoder structure, ShelfNet has multiple encoder-decoderbranch pairs with skip connections at each spatial level, which looks like ashelf with multiple columns. The shelf-shaped structure provides multiple pathsfor information flow and improves segmentation accuracy. Inspired by thesuccess of recurrent convolutional neural networks, we use modified residualblocks where two convolutional layers share weights. The shared-weight blockenables efficient feature extraction and model size reduction. We testedShelfNet with ResNet50 and ResNet101 as the backbone respectively: theyachieved 59 FPS and 42 FPS respectively on a GTX 1080Ti GPU with a 512x512input image. ShelfNet achieved high accuracy: on PASCAL VOC 2012 test set, itachieved 84.2% mIoU with ResNet101 backbone and 82.8% mIoU with ResNet50backbone; it achieved 75.8% mIoU with ResNet50 backbone on Cityscapes dataset.ShelfNet achieved both higher mIoU and faster inference speed compared withstate-of-the-art real-time semantic segmentation models. We provide theimplementation https://github.com/juntang-zhuang/ShelfNet.