01 / INPUTPNG or JPEG / up to 8 MiB
MACHINE LEARNING DEMONSTRATION
Cityscapes segmentation
models and results
Compare U-Net, BiSeNetV2, and PIDNet-S on Cityscapes. Run the selected PIDNet-S checkpoint on an urban road image.
MODELS3 architecturesU-Net / BiSeNetV2 / PIDNet-S
MEASUREMENTSmIoU / throughput / memory
Checking model availability...
Segment an image
View the original image, predicted classes, and color overlay.
02 / PREDICTION
Your scene, segmentedThe prediction will appear here.
Choose an image to begin.
Classes in this prediction
Colors match the labels in the image.
Semantic class colorsNames and descriptions for all 19 Cityscapes labels.
The image is sent to the configured server for inference. Saved predictions are stored in this browser. Accuracy requires a labeled reference image.
Processed images
Earlier uploads, predictions, and their detected classes, saved in this browser.
Loading saved predictions...
No saved predictions yet. Run segmentation on an image to add it here.
History stays in this browser and is removed if you clear it here or clear browser site data.
Cityscapes experiment results
Three completed validation runs. Select a model to inspect its configuration and epoch history.
| Model / run | Configuration | mIoU | Macro Dice/F1 | Epoch time | Images / s | GPU memory |
|---|
No completed research runs found at the configured results path.
Run configuration and history
Official Cityscapes validation split resized to 256 x 512; one seed and three epochs. These are preliminary validation results, not held-out test or leaderboard scores. GPU memory is peak allocated training memory.
Method and provenance
Prepare
Official Cityscapes train and validation splits. Nineteen evaluation classes, paired image and mask transforms, and ignored void labels.
Train and measure
U-Net, BiSeNetV2, and PIDNet-S trained from scratch on a V100 using matched C1 FP32 settings: 256 x 512, batch size 2, seed 42, and three epochs.
Evaluate and demonstrate
Compare validation mIoU, Dice/F1, throughput, and GPU memory. The selected PIDNet-S checkpoint serves CPU image-upload inference.
Active checkpoint
No checkpoint configured.
This is a non-commercial academic demo. U-Net is a documented baseline variant; BiSeNetV2 and PIDNet use attributed upstream implementations. Live CPU inference time is separate from V100 training throughput.