EfficientSeg: An Efficient Semantic Segmentation Network
This work addresses the challenge of efficient training for semantic segmentation in scenarios with few data and no pre-trained weights, offering a scalable solution that improves performance over baselines.
The paper tackles the problem of training deep semantic segmentation networks efficiently without pre-trained weights and limited data, achieving a 58.1% mIoU score on the Minicity dataset and placing fourth in the ECCV 2020 VIPriors challenge.
Deep neural network training without pre-trained weights and few data is shown to need more training iterations. It is also known that, deeper models are more successful than their shallow counterparts for semantic segmentation task. Thus, we introduce EfficientSeg architecture, a modified and scalable version of U-Net, which can be efficiently trained despite its depth. We evaluated EfficientSeg architecture on Minicity dataset and outperformed U-Net baseline score (40% mIoU) using the same parameter count (51.5% mIoU). Our most successful model obtained 58.1% mIoU score and got the fourth place in semantic segmentation track of ECCV 2020 VIPriors challenge.