Computer Vision News - June 2020

2 Summary Research 18 The model took approximately 2 days to train on 2 TitanX gpus. As we were particularly interested on time performance since it looks like a key element for this application, we asked Nicu Sebe about it. “The training time is rather long (with more GPUs it can still be reduced substantially) […], but this is not a real problem as the training is done once and all the process is done offline. What is really important is that the testing is real-time as this is what the user gets to see. Due to the available hardware the approach works fine with small/low-resolution images and videos so the hope is that in the near future we can achieve excellent performance also in creating high-resolution content. ” Conclusion The first order motion model represents an optimal method to obtain realistic image animation for any object belonging to a specific class used to train the model. It does this by wisely combining encoder-decoder, dense motion and generator networks. It also introduces new elements to the field, such as the occlusion mask and the Tai-Chi-HD dataset that can be seen both as results and techniques to evaluate and benchmark the already existent models and the future ones. To dive deeper in this topic, the paper presents an appendix with further results, and it inspired us so much that we decided to also ask some other questions to Nicu about the future directions of his work. content and relax the constraint of needing a driving video. On this topic, he comments: “It would be better if we can allow the creation of videos that for example illustrate the animation of the character in any plausible way, e.g., I can indicate only some attributes like fast/slow/jumpy and be also able to represent this movement in a realistic way, i.e., indistinguishable from a real video .” When asked about the application of this technology, Nicu replied that the “goal for the future is to support users in their attempt to achieve high-quality personalized content generation”. He also mentioned that famous companies such as Facebook and Snap are already moving in this domain through their filters, investigating its potential. In the future, he would like to see an optimization in performance for high-resolution