Segmentation Demo

How to use this demo

All you need to do is to upload an rgb image. If any of its dimensions exceeds 500 pixels, it will be resized to 500.

After uploading the image and choosing appropriate parameters, please click the 'run the model' button. Have fun!

Model description

The current semantic segmentation model, global deconvolutional network, is an adaptation of Deep-Lab LargeFOV ("Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs" by L.C.Chen et al).
In contrast to Deep-Lab, our model relies on global interpolation to upscale the coarse output from the network. Furthermore, the model is also trained with an additional multi-label classification loss to enforce correct recognition of classes.
More details on our model can be found in the paper "Global Deconvolutional Networks for Semantic Segmentation".

CRF description

CRF is a standard tool for semantic image segmentation task. This particular denseCRF is described fully in the paper "Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials" by P. Krähenbühl and V. Koltun.

PASCAL VOC 2012 test results

The model currently shows 74.02% mean IoU accuracy on the test set of the PASCAL VOC benchmark.