ICCV Daily 2021 - Friday

The call for papers expressively asked potential authors to send works that addressed any kind of topic or problem. The only rule was that the paper should not be focused solely on deep learning. That meant researchers could send papers that do not use deep learning at all, or something that mixes deep learning with traditional knowledge or hand-crafted algorithms that could give some new insight into bridging the two worlds. They were keen to offer a place that highlighted works that are hard to find among the numerous deep learning papers on offer. Matteo points out one paper which explores view-graph construction with uncalibrated cameras . “ This is something that you could find in one of the top conferences, ” he tells us. “ For instance, Federica has a paper about a similar topic at the main conference. But you would find maybe one or two among the 1,000 papers at ICCV. There are a couple of papers that show very interesting results against deep approaches. There’s one paper which is more like a survey because they made several experiments about point cloud clustering. They show that traditional clustering approaches for panoptic segmentation work extremely well despite deep learning being the most popular approach nowadays to tackle the problem. ” The workshop has a remarkable line-up of speakers, including Venu Madhav 17 DAILY ICCV Friday Traditional Computer Vision in the Age of Deep Learning Comparison between 3D reconstructions by a deep network and the method proposed in: Finite Aperture Stereo: 3D Reconstruction of Macro-Scale Scenes