Efficient Semantic Segmentation Using Gradual Grouping

Nikitha Vallurupalli, Sriharsha Annamaneni, Girish Varma, C.V. Jawahar, Manu Mathew, Soyeb Nagori; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 598-606

Abstract


Deep CNNs for semantic segmentation have high mem- ory and run time requirements. Various approaches have been proposed to make CNNs efficient like grouped, shuf- fled, depth-wise separable convolutions. We study the ef- fectiveness of these techniques on a real-time semantic segmentation architecture like ERFNet for improving run- time by over 5X. We apply these techniques to CNN lay- ers partially or fully and evaluate the testing accuracies on Cityscapes dataset. We obtain accuracy vs parame- ters/FLOPs trade offs, giving accuracy scores for models that can run under specified runtime budgets. We further propose a novel training procedure which starts out with a dense convolution but gradually evolves towards a grouped convolution. We show that our proposed training method and efficient architecture design can im- prove accuracies by over 8% with depthwise separable con- volutions applied on the encoder of ERFNet and attaching a light weight decoder. This results in a model which has a 5X improvement in FLOPs while only suffering a 4% degra- dation in accuracy with respect to ERFNet.

Related Material


[pdf] [arXiv]
[bibtex]
@InProceedings{Vallurupalli_2018_CVPR_Workshops,
author = {Vallurupalli, Nikitha and Annamaneni, Sriharsha and Varma, Girish and Jawahar, C.V. and Mathew, Manu and Nagori, Soyeb},
title = {Efficient Semantic Segmentation Using Gradual Grouping},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2018}
}