Method Details


Details for method 'BiSeNetV2-CityLabeler'

 

Method overview

name BiSeNetV2-CityLabeler
challenge pixel-level semantic labeling
details This method utilizes the Bilateral Segmentation Network V2 (BiSeNetV2) architecture, which balances real-time inference speed with segmentation accuracy. The model was trained on the Cityscapes dataset using the standard train set with data augmentation techniques including random scaling and cropping. The implementation employs a two-pathway design featuring a Detail Branch that preserves spatial details and a Semantic Branch that captures high-level semantic context. The network uses an auxiliary training strategy with multiple supervision branches to enhance feature representation learning. Inference is performed at the original image resolution (2048×1024) to maintain spatial fidelity, and the model outputs predictions using the Cityscapes labelIDs encoding scheme. The final segmentation results are produced by directly applying the trained model to the test images, without additional post-processing or ensemble methods, allowing for efficient deployment without sacrificing accuracy.
publication Anonymous
project page / code
used Cityscapes data fine annotations
used external data ImageNet
runtime 15 s
CPU:i9-14900K; RTX 3090 24G
subsampling no
submission date April, 2025
previous submissions

 

Average results

Metric Value
IoU Classes 63.0004
iIoU Classes 39.0026
IoU Categories 86.2491
iIoU Categories 73.0547

 

Class results

Class IoU iIoU
road 97.2546 -
sidewalk 77.1308 -
building 89.2508 -
wall 35.9817 -
fence 40.3434 -
pole 50.5546 -
traffic light 57.1955 -
traffic sign 65.7037 -
vegetation 91.6334 -
terrain 67.0706 -
sky 93.432 -
person 78.5198 59.2238
rider 51.4497 29.0189
car 92.4125 86.2746
truck 31.5612 14.5534
bus 36.307 25.3762
train 32.0339 18.4654
motorcycle 41.7898 25.4101
bicycle 67.3822 53.6984

 

Category results

Category IoU iIoU
flat 98.2455 -
nature 91.2824 -
object 58.0207 -
sky 93.432 -
construction 89.8576 -
human 80.7801 61.3801
vehicle 92.1257 84.7294

 

Links

Download results as .csv file

Benchmark page