The 2020 IEEE Low-Power Computer Vision Challenge (LPCVC) concluded successfully on 2020/07/31, after a month of competition. Forty six teams submitted 378 solutions in four tracks. During the month-long challenge, many teams submitted multiple solutions and witnessed the scores becoming much higher. The 2020 LPCVC adds a video track and expands the Low-Power Image Recognition Challenge (LPIRC) started in 2015.

The 2020 LPCVC has three tracks:

1. Recognizing English letters in video taken by an unmanned aerial vehicle (UAV, also called drone) using PyTorch Mobile
2. Recognizing objects in images using FPGA (Field Programming Gate Array)
3. Recognizing objects in images using TensorFlow
4. Classifying images using TensorFlow

For more information, go to the competition introduction page. To see a detailed score breakdown for each submission, go to the leaderboard.

PyTorch - Video Track

This track requires that participating teams recognize English letters or numbers in the signs in the video captured by an unmanned aerial vehicle (UAV). The following two video frames show examples of such a challenge. The challenge appears in the form of question-answer pairs. A successful solution has to answer the English letters or numbers appearing in the same frame as the questions. For example, if the question is “RESTRICTED AREA”, the answer should be “B056 ELEVATOR CONTROL ELEVATOR PERSONNEL ONLY”. If the question is “CONFERENCE”, the answer should be “122”. The solutions are case insensitive. The solutions must use PyTorch running on Raspberry Pi 3B+.

Eleven teams submitted 84 solutions to this track. The winners of this tracks are

FPGA Track

This track uses Ultra96-V2 + Xilinx® Deep Learning Processor Unit (DPU) running software PYNQ + Vitis AI. The challenge recognizes objects in images. Nineteen teams submitted 115 solutions. The winners of this track are:

TensorFlow Track (also called On-Device Visual Intelligence Challenge, OVIC)

This OVIC track contains three tracks:

1. Interactive object detection: This category focuses on COCO detection models operating at 30 ms / image on a Pixel 4 smartphone (CPU).
2. Real-time image classification: This category focuses on Imagenet classification models operating at 10 ms / image on a Pixel 4 smartphone (CPU).
3. Real-time image classification: This category focuses on Imagenet classification models operating at 7 ms / image on a LG G8 smartphone (DSP).

Sixteen teams submitted 179 solutions to these tracks.

The winners of the first track (Interactive object detection using Pixel 4 CPU) are:

The winners of the second track (Real-time image classification using Pixel 4 CPU) are:

The winners of the third track (Real-time image classification using LG G8 DSP) are:

LPCVC started in 2015 with the aim to identify the technologies for computer vision using energy efficiently. IEEE Rebooting Computing has been the primary financial sponsor and provides administrative support. In 2020, sponsors are PyTorch, Google, Mediatek, Xilinx, IEEE Council of Design Automation, ELAN Microelectronics, National Science Foundation, IEEE Circuits and Systems Society, IEEE Rebooting Computing. A new IEEE technical committee has been created to coordinate future activities of low-power computer vision. People interested in joining the committee please contact Terence Martinez, Program Director Future Directions IEEE Technical Activities, t.c.martinez@ieee.org.

The organizing committee of the 2020 Low-Power Computer Vision Challenge are