Computer Vision News 6 data for training. There may also be ways to improve its effectiveness. “We propose a way to make the synthetic data we have at hand more effective in training the classifier by introducing some prior information from pretrained models,” she adds. “Then, by using it to regularize the network, it improves the performance of the synthetic data without changing the generative model itself.” Reflecting on her journey, Ruyu tells us the initial spark that ignited her interest in computer science was her childhood love for computer games. However, realizing that game design was a lot about physics and mathematics led her to explore other directions. It was during her master’s thesis in a lab focused on traffic-related applications that she first encountered anomaly detection, particularly in identifying irregularities on road surfaces. “At that time, I really thought I was crazy, and my professor was crazy because he gave me the money to buy a drone!” she laughs. “I flew it to capture an aerial traffic video to compute traffic from the video. That’s where I started to touch this deep learning and computer vision thing. It all started with Faster R-CNN in 2015. I used it to do the car detection. Then I used a Kalman filter-based algorithm to trace the cars and do the counting.” Ruyu came to Germany for her exchange semester, where, compared to Taiwan, which is more focused on semiconductor production, she recognized the potential for a future in computer vision and autonomous driving or different kinds of production. Wanting to stay, she sought opportunities for further exploration, leading to an internship at Bosch. “That’s where I got into generative models because my supervisor’s idea at that time was to use CycleGAN to transfer an okay product to a defective product,” she recalls. “I just thought, these generative model things look pretty cool! My supervisor also said, ‘Don’t you think working on images is way cooler than other things because you see the images you produce?’ I agree. It’s very interesting. That’s why I’ve kept doing it until now!” WACV Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=