How much does the batching help? Small benchmark of 8 images, 300 queries, 10 targets/image Sequential: 7.66 ms Grouped: 1.26 ms -> 6.1× faster matcher execution (Both paths used the CUDA solver.…

That cost typically combines: * classification cost * L1 box distance * IoU / GIoU Then many implementations call SciPy’s linear_sum_assignment(). If the model is training on GPU, that means the…