Computer Vision News 24 things that did not work out and realized the depth sensor was super important for initialization and resampling.” The breakthrough came when integrating monocular depth and normal priors, enabling the system to complete extensive sequences and significantly improving over previous attempts that often failed early in the process. “We always failed in the first 200-400 frames, which was really frustrating,” Zihan recalls. “As soon as we tried this monocular and normal depth, it improved, and you can at least finish the whole 2,000-frame sequence.” Further enhancements were achieved by incorporating warping loss and optical flow loss, bolstering the system’s accuracy. Following a well-received presentation at 3DV by Zihan, the team took home a Best Paper Honourable Mention award for their efforts, a recognition that caught both authors by surprise and one that came at the end of a long road that began with rejections from SIGGRAPH and SIGGRAPH Asia last year. “Most of the reviews were nice, but they didn’t think SLAM was suitable for the SIGGRAPH community,” Zihan reveals. What do they think the judges at 3DV saw in the work? “If I try to think about why we got this award, I’d say it’s because we have good visualizations,” Songyou asserts confidently. “We show really good results in novel view synthesis and surface rendering and decent results on tracking as well. Also, we provide a solution for a challenging task. It might not be the fastest, but we provide a clue that it’s possible.” Best Paper 3DV Hon. Mention
RkJQdWJsaXNoZXIy NTc3NzU=