aa200647963/sgg-dhl

This repository contains code for the paper 'Dual-branch Hybrid Learning Network for Unbiased Scene Graph Generation'.

sgg-dhl's Issues

about the new version of your paper?

i see your new version paper. there is caption experiments.
Not sure if you have any plans to open source code for this part, I'm doing similar research recently and hope to get some inspiration from your research. Many thanks for this.

about the hyperreference settings of vctree and motif?

hi, dear sir.
can you update the scripts of vctree and motif?
i just use the same hyperreference settings of transformer with vctree model, but the precision is bad.
so hope you can provide the scripts.

pre-trained object detector model

Hello author, I have seen that the experiment of GQA dataset has been added to the paper, I wonder if you can provide a pre-trained object detector model for this dataset, thank you.

About the installation of maskrcnn_benchmark

Hi, I have a question about the installation. In INSTALL.md you install maskrcnn_benchmark from https://github.com/KaihuaTang/Scene-Graph-Benchmark.pytorch.git, while in the repository you provide a directory of maskrcnn_benchmark, which seems to be different. I wonder which one should I install then?

the experimental results are not as good as the results presented in your paper?

Dear sir:
I repeated your open source transformer experiment, but found that the experimental results are not as good as the results presented in your paper and log files, its mr performance is excellent, but r performance is not good. Is there a problem of large fluctuations in this training process, because I see that your loss has been dynamically adjusted. Do you have any adjustment suggestions?

log.txt

about the SGDet log.txt?

I just modify the predcls script to sgdet type with transformer model, but the result is lower 2 points on R and 2 points on mR, can you upload the scripts and log.txt of the sgdet model?

about the training speed?

hi, i notice speed of your code is faster than Tang's original by twice times. i just want to know how do you do that, i dont find this.

about the train log?

hi, can you provide the model or log file of your model? thank you very much!