search

YiwuZhong / SGG_from_NLS

[ICCV 2021] Official code for "Learning to Generate Scene Graph from Natural Language Supervision"
Other
100 stars 10 forks source link

use the code for custom images #3

Closed DeepaliVerma closed 2 years ago

DeepaliVerma commented 2 years ago

Sir, how can I use this code to predict scene graphs on random images rather than the dataset images.

YiwuZhong commented 2 years ago

Thanks for your interest in our paper! To run on custom images, before using the SGG models, the first step is to detect objects and extract their features.

In most of our experiments, following VSPNet, we used the object detector trained on OpenImage-V4 from Tensorflow. This post might be useful for region feature extraction. Or the other way, you can use the object detector trained on Visual Genome. This detector was used in our fully supervised experiment and can be combined with our trained model Sup_OnlineDetector_Uniter in the README. The command can be referred to SGDet on Custom Images.

YiwuZhong commented 2 years ago

There is a new update about extracting region features using Tensorflow Object Detection API. If you're interested in using Tensorflow API, the discussion in post #5 might be helpful.