Real-time Semantic Segmentation of Urban Scenes on the Cityscapes Dataset Using Deeplabv3 (resnet50) With Web Deployment
Year: 2026
Type: article
Abstract: Autonomous driving and robotics systems can't get very far with a single label slapped on an entire image -they need to know what every pixel in front of them actually is.That is the problem this work tackles, using the Cityscapes dataset, a collection of high-resolution street photographs from 50 European cities, to build a real-time semantic segmentation pipeline.We fine-tune a pretrained DeepLabV3 network with a ResNet50 backbone that assigns each pixel of a street image to one of 19 standard classes -road, building, person, car, vegetation, sky, traffic light, pole, and so on.To keep training and inference feasible on limited hardware, the input images are downsized from 2048×1024 to 512×256.Atrous (dilated) convolutions and an Atrous Spatial Pyramid Pooling (ASPP) module widen the receptive field and capture multi-scale context without any extra downsampling.Because classes like road and building vastly outnumber rare ones like poles and traffic lights, the loss function is class-weighted to counter this imbalance.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7237355603
