Arbitrary object detection with transformers
Testing out some ideas in feature injection in the convolutional backbone before the attention head for the DETR architecture. Mainly to speed up convergence and minimize reliance on global attention.
This was before I read the DAB-DETR papers, amongst the multitudes of variants of DETR out there. Each variant seems more convoluted and arbitrary than the last...
More thought needs to be done in this domain, namely that of motivating design choices via some kind of unified framework, similar to geometric deep-learning. People are using category theory nowadays.