Use Case and High-Level Description

A trained model of ICNet for fast semantic segmentation, trained on the CamVid* dataset from scratch using the TensorFlow* framework. The trained model has 30% sparsity (ratio of zeros within all the convolution kernel weights). For details about the original floating-point model, check out the ICNet for Real-Time Semantic Segmentation on High-Resolution Images.

The model input is a blob that consists of a single image of 1x3x720x960 in the BGR order. The pixel values are integers in the [0, 255] range.

The model output for icnet-camvid-ava-sparse-30-0001 is the predicted class index of each input pixel belonging to one of the 12 classes of the CamVid dataset:

  • Sky
  • Building
  • Pole
  • Road
  • Pavement
  • Tree
  • SignSymbol
  • Fence
  • Vehicle
  • Pedestrian
  • Bike
  • Unlabeled


Metric Value
GFlops 151.82Bn
MParams 25.45
Source framework TensorFlow*


The quality metrics were calculated on the CamVid* validation dataset. The unlabeled class had been ignored during metrics calculation.

Metric Value
mIoU 69.99%
  • IOU=TP/(TP+FN+FP), where:
    • TP - number of true positive pixels for given class
    • FN - number of false negative pixels for given class
    • FP - number of false positive pixels for given class


Image, shape - 1,3,720,960, format is B,C,H,W where:

  • B - batch size
  • C - channel
  • H - height
  • W - width

Channel order is BGR.


Semantic segmentation class prediction map, shape - 1,720,960, output data format is B,H,W where:

  • B - batch size
  • H - horizontal coordinate of the input pixel
  • W - vertical coordinate of the input pixel

Output contains the class prediction result of each pixel.

Legal Information

[*] Other names and brands may be claimed as the property of others.