This paper introduces a fully convolutional DenseNet architecture for semantic segmentation, called FC-DenseNet. The proposed architecture extends the DenseNet, which is known for its efficient parameter usage and deep supervision, to handle semantic segmentation tasks. The key idea is to add an upsampling path to recover the full input resolution, which allows the network to perform semantic segmentation without any post-processing module or pretraining. The architecture is designed to avoid the feature map explosion that occurs in naive extensions of DenseNets, by only upsampling the feature maps created by the preceding dense block. This approach results in a model with significantly fewer parameters than current state-of-the-art models for urban scene understanding datasets such as CamVid and Gatech. The model achieves state-of-the-art results on these datasets without any additional post-processing or pretraining. The proposed architecture is evaluated on two challenging benchmarks for urban scene understanding, CamVid and Gatech, and shows significant improvements in performance compared to existing methods. The model is trained from scratch and achieves high accuracy on both datasets. The results demonstrate that the proposed architecture is effective for semantic segmentation and can outperform current state-of-the-art results without any additional post-processing or pretraining. The paper also discusses the advantages of the proposed architecture, including parameter efficiency, implicit deep supervision, and feature reuse. The model is able to outperform state-of-the-art methods without requiring any temporal smoothing, and it is shown that the model can benefit from pretraining on large datasets such as ImageNet. The paper concludes that the proposed architecture is a promising approach for semantic segmentation and has the potential to be applied to a wide range of tasks.This paper introduces a fully convolutional DenseNet architecture for semantic segmentation, called FC-DenseNet. The proposed architecture extends the DenseNet, which is known for its efficient parameter usage and deep supervision, to handle semantic segmentation tasks. The key idea is to add an upsampling path to recover the full input resolution, which allows the network to perform semantic segmentation without any post-processing module or pretraining. The architecture is designed to avoid the feature map explosion that occurs in naive extensions of DenseNets, by only upsampling the feature maps created by the preceding dense block. This approach results in a model with significantly fewer parameters than current state-of-the-art models for urban scene understanding datasets such as CamVid and Gatech. The model achieves state-of-the-art results on these datasets without any additional post-processing or pretraining. The proposed architecture is evaluated on two challenging benchmarks for urban scene understanding, CamVid and Gatech, and shows significant improvements in performance compared to existing methods. The model is trained from scratch and achieves high accuracy on both datasets. The results demonstrate that the proposed architecture is effective for semantic segmentation and can outperform current state-of-the-art results without any additional post-processing or pretraining. The paper also discusses the advantages of the proposed architecture, including parameter efficiency, implicit deep supervision, and feature reuse. The model is able to outperform state-of-the-art methods without requiring any temporal smoothing, and it is shown that the model can benefit from pretraining on large datasets such as ImageNet. The paper concludes that the proposed architecture is a promising approach for semantic segmentation and has the potential to be applied to a wide range of tasks.