Diana Wofk—Fast and energy-efficient monocular depth estimation on embedded systems

Описание: Diana Wofk, a recent Masters in Engineering graduate from the Department of Electrical Engineering & Computer Science (EECS), gave the Nano Explorations talk on Thursday, July 23, 2020.

Depth sensing is a critical function for many robotic tasks such as localization, mapping and obstacle detection. There has been a significant and growing interest in performing depth estimation from monocular RGB camera images, due to the relatively low cost, size, weight and power of cameras. However, state-of-the-art depth estimation algorithms are based on fairly large deep neural networks, which have high computational complexity and energy consumption. This poses a significant challenge when performing real-time depth estimation on an embedded system, for instance, a mobile phone or a platform mounted on a Micro Aerial Vehicle (MAV).

Our work addresses this problem of fast and energy-efficient depth estimation on embedded platforms. Our proposed network, FastDepth, runs at 178 fps on the Jetson TX2 embedded GPU, with active power consumption of 8.8 W. We seek to further improve energy efficiency by deploying onto a low-power embedded FPGA. Using an algorithm-hardware co-design approach, we develop a dataflow design and an accelerator architecture that minimizes off-chip memory accesses and offers dedicated support for depthwise separable convolutional layers. This talk will give an overview of our approach and the strategies we take in accelerating learning-based depth estimation on embedded systems.

---------------------------------------------------------

Nano Explorations is a virtual seminar series powered by MIT.nano that features presentations by MIT students on their work in nanoscience, nanotechnology, and other advanced research fields. The series was launched as a way to keep our nano community connected during the COVID-19 pandemic. Presentations take place on Tuesdays and Thursdays at 11am EST and are open to any interested viewers.

Read more about past and upcoming Nano Explorations at mitnano.mit.edu/nano-explorations.