Computer Vision News - May 2024

3 Computer Vision News Previously, works on this topic commonly modeled uncertainty by treating the predicted poses and shapes as probability distributions, calculating confidence by sampling multiple instances of the predicted distributions, and then analyzing the variability or deviation among these samples. “However, this approach has a problem,” Sai tells us. “First, it’s very slow because you have to do multiple samples to get the uncertainty estimate. Second, there’s a trade-off between speed and accuracy.” POCO can convert any state-of-theart human pose and shape regressor into an approach that estimates the uncertainty quickly and accurately in a single forward pass. “The important thing is it doesn’t hurt the accuracy of the pose estimation,” Dimitrios confirms. “POCO gives you an extra modality as output, which is the confidence of the estimation, but the accuracy of the pose estimation itself is not compromised.” In fact, in a recent LinkedIn post, Michael Black revealed that POCO improves the method’s accuracy. “An interesting byproduct is that we find that training methods to estimate their certainty actually makes them more accurate – the effect is small but consistent across all regressors tested.” He added: “There is only upside to using POCO.” Nevertheless, the journey toward integrating confidence with pose estimation has not been without its challenges. Supervising uncertainty where there is no ground truth is very complex, and convincing people that it is possible to use it without affecting performance can be difficult. “The computer vision community is laser-focused on benchmarks,” Dimitrios points out. “We want accuracy to be better, but sometimes we forget about the useful context around it. We want our methods to be very accurate, but we also need to know when to trust them!” POCO