Closed hutao568 closed 3 years ago
I am sorry that this checkpoint is missing, (only DLA-34 is kept) and I will run it and provide the checkpoint if I have spare time (Maybe in June). The provided accuracy for HOI-A in the paper was achieved by the old-version codes, where the human branch and object branch are split. During training, we modify the target generation (Gaussian redius) and loss weight for the small objects, especially smoke. You also can take our new method AS-Net as a reference, which has achieved higher performance in HOI-A.
Hello, I used the code you provided to fail to achieve the accuracy in your paper (71.23@mAP) on the HOIA dataset. I am using the hourglass model. Can you provide checkpoints?