algorithm [ R1

We thank reviewers for the constructive comments. We will release all models and further polish the documents. The result on COCO val set is 44.0 PQ (+0.1 PQ vs no inter-modular, -0.7 PQ vs Auto-Panoptic). DF conv in both heads achieves 44.8 PQ, which is 0.4 PQ lower and is much slower due to the heavy head. Hand-crafted model in Table 4 clearly demonstrates the effectiveness of the search algorithm.