ApproxNet: Content-Aware Approximation for Video Classification on Mobile Platforms

This is a demo of ApproxNet. It shows a sample video downloaded from YouTube on which we are running an object classification task using an approximate version of a ResNet. This version is lightweight enough that it can run on a mobile device and can keep up with 30 fps. For comparison, we show the previous state-of-the-art solution, MCDNN [Han et al. Mobisys 2016].

ApproxNet: Content and Contention-Aware Video Object Classification System for Embedded Clients,” Resilient Wireless Networks
Ran Xu, Rakesh Kumar, Pengcheng Wang, Peter Bai, Ganga Meghanath (IIT Madras), Somali Chaterji, Subrata Mitra (Adobe Research), and Saurabh Bagchi. Accepted to appear in ACM Transactions on Sensor Networks (TOSN), pp. 1-27, 2021.

The figure depicts the class predictions of ApproxNet along with latency and accuracy plots comparing the performance of ApproxNet and MCDNN for each frame of the video. We can see that ApproxNet is always faster than MCDNN without any switching overhead, while MCDNN incurs significant switching overhead when changing model variants (as it has to do when the video characteristics change significantly enough).

Predictions by ApproxNet vs MCDNN for a car, over a timeline
Predictions by ApproxNet vs MCDNN for an airplane, over a timeline
Predictions by ApproxNet vs MCDNN for an animal, over a timeline
Last modified: May 5, 2021