The first part of this page presents the results for single-task results. This means that we care about the maximum performance a model can reach on each task, disregarding the other ones. The second part of the page presents the methods which, using a single model (i.e. a single set of weights for a neural network or a self-contained single machine learning algorithm) it predicts all three tasks at once with a given performance.
For the exact formulas needed to compute these metrics with your own model, see the end of the page.
The first part of this page presents the results for multi-task results. The results here are provided through a single model which predicts all three tasks at once (i.e. a single set of weights for a neural network or a self-contained single machine learning algorithm). We will soon update this section to include a way of combining the three metrics into a single non-weighted number so each task contributes the same.
For the real time models, the authors need to provide a self-contained Google Colab notebook. The models must run at >=5FPS on a T4 GPU (free Google Colab instance) for the Dronescapes-Test (or similar) images at 960x540 resolution. We provide an example for the PHG-MAE-Distil 430k parameters model: link.
TODO