JejuNet

Real-Time Video Segmentation on Mobile Devices

Keywords

Video Segmentation, Mobile, Tensorflow Lite

Tutorials

Benchmarks: Tensorflow Lite on GPU
- A Post on Medium Link
- Detail results Link

Introduction

Running vision tasks such as object detection, segmentation in real time on mobile devices. Our goal is to implement video segmentation in real time at least 24 fps on Google Pixel 2. We use efficient deep learning network specialized in mobile/embedded devices and exploit data redundancy between consecutive frames to reduce unaffordable computational cost. Moreover, the network can be optimized with 8-bits quantization provided by tf-lite.

Example: Reai-Time Video Segmentation(Credit: Google AI)

Architecture

Video Segmentation

Compressed DeepLabv3+[1]
- Backbone: MobileNetv2[2]

Optimization

8-bits Quantization on TensorFlow Lite

Experiments

Video Segmentation on Google Pixel 2
Datasets
- PASCAL VOC 2012

Plan @Deep Learning Camp Jeju 2018

July, 2018

DeepLabv3+ on tf-lite
Use data redundancy between frames
Optimization
- Quantization
- Reduce the number of layers, filters and input size

Results

More results here bit.ly/jejunet-output

Demo

Video Segmentation on Google Pixel 2

Trade-off Between Speed(FPS) and Accuracy(mIoU)

Low Bits Quantization

Network	Input	Stride	Quantization(w/a)	PASCAL mIoU	Runtime(.tflite)	File Size(.tflite)
DeepLabv3, MobileNetv2	512x512	16	32/32	79.9%	862ms	8.5MB
DeepLabv3, MobileNetv2	512x512	16	8/8	79.2%	451ms	2.2MB
DeepLabv3, MobileNetv2	512x512	16	6/6	70.7%	-	-
DeepLabv3, MobileNetv2	512x512	16	6/4	30.3%	-	-

References

Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation

Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, Hartwig Adam. arXiv: 1802.02611.

[link]. arXiv: 1802.02611, 2018.
Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation
Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
[link]. arXiv:1801.04381, 2018.

Authors

Taekmin Kim(Mentee) @tantara
Jisung Kim(Mentor) @runhani

Acknowledgement

This work was partially supported by Deep Learning Jeju Camp and sponsors such as Google, SK Telecom. Thank you for the generous support for TPU and Google Pixel 2, and thank Hyungsuk and all the mentees for tensorflow impelmentations and useful discussions.

tantara/JejuNet