This repository contains the source code of our paper, ESPNet (accepted for publication in ECCV'18).

Check our project page for more qualitative results (videos).

Click on the below sample image to view the segmentation results on YouTube.

This repository is organized as:

• train This directory contains the source code for trainig the ESPNet-C and ESPNet models.
• test This directory contains the source code for evaluating our model on RGB Images.
• pretrained This directory contains the pre-trained models on the CityScape dataset
  • encoder This directory contains the pretrained ESPNet-C models
  • decoder This directory contains the pretrained ESPNet models

Our model ESPNet achives an class-wise mIOU of 60.336 and category-wise mIOU of 82.178 on the CityScapes test dataset and runs at

• 112 fps on the NVIDIA TitanX (30 fps faster than ENet)
• 9 FPS on TX2
• With the same number of parameters as ENet, our model is 2% more accurate

Our model achieves an mIOU of 55.64 on the CamVid test set. We used the dataset splits (train/val/test) provided here. We trained the models at a resolution of 480x360. For comparison with other models, see SegNet paper.

Note: We did not use the 3.5K dataset for training which was used in the SegNet paper.

To run this code, you need to have following libraries:

• OpenCV - We tested our code with version > 3.0.
• PyTorch - We tested with v0.3.0
• Python - We tested our code with Pythonv3. If you are using Python v2, please feel free to make necessary changes to the code.

We recommend to use Anaconda. We have tested our code on Ubuntu 16.04.

If ESPNet is useful for your research, then please cite our paper.

@inproceedings{mehta2018espnet,
  title={ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation},
  author={Sachin Mehta, Mohammad Rastegari, Anat Caspi, Linda Shapiro, and Hannaneh Hajishirzi},
  booktitle={ECCV},
  year={2018}
}

If you are getting an assertion error with class labels, then please check the number of class labels defined in the label images. You can do this as:

import cv2
import numpy as np
labelImg = cv2.imread(<label_filename.png>, 0)
unique_val_arr = np.unique(labelImg)
print(unique_val_arr)

The values inside unique_val_arr should be between 0 and total number of classes in the dataset. If this is not the case, then pre-process your label images. For example, if the label iamge contains 255 as a value, then you can ignore these values by mapping it to an undefined or background class as:

labelImg[labelImg == 255] = <undefined class id>