Thank you Ross! We often use these frameworks and even conduct research on top of them without taking a step back and considering that they are all based on a ML-heavy modeling of the problems and in the quantization/regression paradigm, which comes with a series of challenges. I was particularly interested in your analysis that some of the most recent frameworks are shifting from geometric entities-based models to object queries that will eventually learn, or be trained to, output geometric entities! Amazing. I think a lot of people will be reading about DETR because of the points you made here. Thanks again and keep up the great work at FAIR.
