Abstract
Recent advances in semantic segmentation generally adapt an ImageNet pretrained backbone with a special context module after it to quickly increase the field-of-view. Although successful, the backbone, in which most of the computation lies, does not have a large enough field-of-view to make the best decisions. Some recent advances tackle this problem by rapidly downsampling the resolution in the backbone while also having one or more parallel branches with higher resolutions. We take a different approach by designing a ResNeXt inspired block structure that uses two parallel 3x3 convolutional layers with different dilation rates to increase the field-of-view while also preserving the local details. By repeating this block structure in the backbone, we do not need to append any special context module after it. In addition, we propose a lightweight decoder that restores local information better than common alternatives. To demonstrate the effectiveness of our approach, our model RegSeg achieves state-of-the-art results on real-time Cityscapes and CamVid datasets. Using a T4 GPU with mixed precision, RegSeg achieves 78.3 mIOU on Cityscapes test set at 30 FPS, and 80.9 mIOU on CamVid test set at 70 FPS, both without ImageNet pretraining.
Abstract (translated)
URL
https://arxiv.org/abs/2111.09957
