Use your
own photo
▶ Load both models & run
Two segmentation networks run
in your browser
(about 45 MB the first time). Your photo is never uploaded.
Overlay
Hover a pixel to see what each model thinks it is.
FCN
ResNet-50 backbone with dilated convolutions (output stride 8) · PASCAL VOC, 21 classes
show the raw 1/8-resolution output (before upsampling)
DeepLab v3
MobileNetV2 backbone, reduced ASPP (1×1 + image pooling) ·
PASCAL VOC, 21 classes
Cityscapes, 19 street classes
ADE20K, 150 classes