Open nicehuster opened 1 month ago
Hi there, Actually, your procedure is correct. The ckpt provided in the repo is from a model trained with further improvements, whereas the ckpt used in our paper was from an earlier version that did not undergo extended training.
it's ok to give some details about the further improvements?
i evaluated the checkpoint at task of referring object classification on LVIS, the reproduced results is so much higher your official paper, your paper(LVIS, sim:87.06, IoU: 62.9 ), the reproduced results(sim:0.927,iou:0.698). Otherwise, the reproduced results on RefCOCOg is (METEOR:28.8, CIDEr: 384.3), your paper(METEOR:23.9, CIDEr: 162.5), where is the problem?