Bayesian SegNet is a deep learning framework for probabilistic pixel-wise semantic segmentation that incorporates model uncertainty. The framework uses Monte Carlo dropout at test time to approximate the posterior distribution of pixel class labels, enabling the model to quantify uncertainty. This approach improves segmentation performance by 2-3% across state-of-the-art architectures like SegNet, FCN, and Dilation Network, without additional parameterization. Bayesian SegNet is benchmarked on the SUN Scene Understanding and CamVid datasets, showing significant improvements, especially on smaller datasets. The model's uncertainty output helps in decision-making by indicating confidence levels in segmentation results. Bayesian SegNet is implemented using the Caffe library and provides a probabilistic segmentation output, with uncertainty measured as the variance of softmax samples. The framework is tested on various benchmarks, including CamVid, SUN RGB-D, and Pascal VOC, demonstrating its effectiveness in semantic segmentation tasks. The model's uncertainty is shown to be a reliable measure of prediction confidence, with higher uncertainty at object boundaries and for visually ambiguous objects. Bayesian SegNet is also applicable to other deep learning architectures, achieving improved segmentation accuracy when applied to FCN and Dilation Network. The model's performance is evaluated using metrics like global accuracy, class average accuracy, and intersection over union. Bayesian SegNet is trained end-to-end with stochastic gradient descent and can run in real-time on a GPU. The framework's ability to model uncertainty makes it a valuable tool for scene understanding tasks, providing reliable predictions and confidence measures for decision-making.Bayesian SegNet is a deep learning framework for probabilistic pixel-wise semantic segmentation that incorporates model uncertainty. The framework uses Monte Carlo dropout at test time to approximate the posterior distribution of pixel class labels, enabling the model to quantify uncertainty. This approach improves segmentation performance by 2-3% across state-of-the-art architectures like SegNet, FCN, and Dilation Network, without additional parameterization. Bayesian SegNet is benchmarked on the SUN Scene Understanding and CamVid datasets, showing significant improvements, especially on smaller datasets. The model's uncertainty output helps in decision-making by indicating confidence levels in segmentation results. Bayesian SegNet is implemented using the Caffe library and provides a probabilistic segmentation output, with uncertainty measured as the variance of softmax samples. The framework is tested on various benchmarks, including CamVid, SUN RGB-D, and Pascal VOC, demonstrating its effectiveness in semantic segmentation tasks. The model's uncertainty is shown to be a reliable measure of prediction confidence, with higher uncertainty at object boundaries and for visually ambiguous objects. Bayesian SegNet is also applicable to other deep learning architectures, achieving improved segmentation accuracy when applied to FCN and Dilation Network. The model's performance is evaluated using metrics like global accuracy, class average accuracy, and intersection over union. Bayesian SegNet is trained end-to-end with stochastic gradient descent and can run in real-time on a GPU. The framework's ability to model uncertainty makes it a valuable tool for scene understanding tasks, providing reliable predictions and confidence measures for decision-making.