Closed thepark96 closed 3 years ago
I realize it's not super well documented. When saving information about a particular object detection, the top-k choices for the predicted object class are saved. That is, the full distribution across all classes isn't preserved just the top 5 indices and their corresponding scores. I don't think I ever actually used this info in practice, but it was helpful just for seeing which classes were often confused. You can find the relevant postprocessing code here: https://github.com/princeton-vl/px2graph/blob/master/task/graph.py#L656
I don't know if you'll see this article, but in the load_prediction function, can you explain me more details about obj.score and rel.score? And is it plausible to think that object class index is the highest index in obj.class_score?