It’s exciting that this repo provides implementation of wide range of methods. However, it’s nearly impossible or at least unclear to me how to apply this implementation to a new dataset.
This repo requires the data to be in a very specific format (ex. bbox distributions, ROI proposals, in the Visual Genome format) because we need to finetune for our new set of classes. To my knowledge, we don't have access to some of these pieces of information on a new dataset.
In light of this: how do people adapt this repo to a new dataset?
An ideal scene graph generation model should be able to trained given input (an image) and output (the object, person, and relationships). I don’t see how a dataset with this annotation can be directly or indirectly applied to this repo’s code.
It’s exciting that this repo provides implementation of wide range of methods. However, it’s nearly impossible or at least unclear to me how to apply this implementation to a new dataset.
This repo requires the data to be in a very specific format (ex. bbox distributions, ROI proposals, in the Visual Genome format) because we need to finetune for our new set of classes. To my knowledge, we don't have access to some of these pieces of information on a new dataset.
In light of this: how do people adapt this repo to a new dataset?
An ideal scene graph generation model should be able to trained given input (an image) and output (the object, person, and relationships). I don’t see how a dataset with this annotation can be directly or indirectly applied to this repo’s code.
Any comment or help are appreciated.