UMEC(HK)’s Summer Interns Win 2nd Runner-up in AI Millionaire – Image Recognition Challenge
UMEC(HK) has been innovating the AI vision technology to bring more convenience and improve the quality of life. Diversified AI applications were developed using our dedicated chips and systems. Three young talents – Shine, Ka Hei and April joined UMEC(HK) as Interns this summer. Apart from learning the latest AI technologies and industry practice, they had the chance to participate in the AI Millionaire – Image Recognition Challenge, organised by Hong Kong Science & Technology Parks Corporation. This had been a precious experience for them to turn knowledge into practice and compete with a number of brilliant teams.
AI Millionaire requires contestants to develop their own deep learning models and train the dataset for the competition. Different AI visual technologies such as simultaneous multi-face detection, face recognition, gender identification and GAN-generated image detection have to be applied to answer the questions about various Asian celebrity photos.
Under the guidance of our experienced AI experts, the interns gained deeper understanding of the advanced AI vision technologies. They worked closely with the mentors to develop and enhance the image recognition system using the modified YOLO v3. Over 300K Asian celebrity photos were collected and labelled for training the AI models. Their endeavours did pay off. Their team, NOVA, won the 2nd runner-up in the competition.
All contestants ran their models on the same virtual machine platform during the competition, but for real-world edge recognition applications, the system performance can be further boosted with customised hardware. Applying network compression, quantisation and other hardware design optimisation techniques, UMEC(HK) further reduces network computation and model size, delivering a more power-efficient hardware solution. Our face recognition ASIC design processes 25-30 fps with power consumption of less than 2 Watts. It supports simultaneous multi-face detection with more than 20 faces in a single photo, achieving over 97% accuracy of face recognition. Furthermore, anti-spoofing 3D face recognition is applied to differentiate between real human faces and flat photos, so as to enhance security.
We are glad to have these young and aspiring talents in our team this summer. With the hands-on experience on machine-learning research, they started a life-long journey of innovations and breakthroughs. We look forward to their future achievements in technology field and welcoming even more ambitious talents to join us in the future!