Closed SISTMrL closed 3 years ago
@SISTMrL Our evaluation showed that pre-training the detector in the HICO-DET dataset and fine-tuning it shows an mAP over 26, but we rather did not include it because of its unnecessarily inefficient training procedure (as we obtained comparable performance only by jointly training the final FFN layer of the detector).
We can include the pre-trained weights in our repo, but it will be during October since me and my resources are currently occupied in preparing our next paper! As our detector follows the exact parameter names and structure of DETR, you can fine-tune it yourself and feed the weights by --resume option!
@meliketoy thanks very much!
Hi,
Can I get the final models hico_q16.pth and vcoco_q16.pth, please ?
hello, have you tried to finetune the whole structure of detr on hicodet datasets later? i want to know the performance gap with detr without finetune.