Yuren Cong

Hi, sry for the late reply. I we used the standard evaluation metrics of OpenImages VRD Challenge. You can find it in the RelDN repo: https://github.com/NVIDIA/ContrastiveLosses4VRD/blob/master/lib/datasets_rel/ap_eval_rel.py

You could have a look at [29](https://github.com/yrcong/STTran/issues/29).

The CUDA version you are using is too high. If you are not using a RTX30 GPU, it is better to downgrade to Cuda10.2.

Hi, sry for the late reply. Actually, we also have no idea. We directly download the dataset from https://www.actiongenome.org/ Although the authors didn't show this category in the [original AG...

Hi, i think it is not difficult to use the model to infer the customized videos. Line 167-178 in https://github.com/yrcong/STTran/blob/main/dataloader/action_genome.py may help you:) best

> If we test on custom video, it seems that the information like attention_relationship and bboxes are required? Why? The attention relationships should be predicted and the bboxes should be...

If you just want to test on your customized video dataset, only the video frames (self.video_list in the class AG) are necessary (for the setting SGDET). Sometimes person_bbox and object_bbox...

Regardless of GPU memory constraint, ResNet101 can be used definitely. You can directly use the pretrained ResNet101 provided by Pytorch, as we used Res50.

Hi, do you mean the pre-processing for the dataset? For different datasets (OpenImages and Visual Genome) we have different scripts. It is somehow complex. I recommend that you follow the...

Which pytorch version did you use? Could you provide more information like the error log?