Learning Everything about Anything: Webly-Supervised Visual Concept Learning

given any concept, discovers an exhaustive vocabulary explaining all its appearance variations

이미지와 컨셉이 주어지면, 그 컨셉 내에서 이미지를 설명하는 단어를 예측한다.

CrissCrossed Caption

USE(Bag-of-Words) / InceptionV3 / ResNet-152 / SimCLRv2 / VSE++ / VSRN

를 사용해서 성능을 측정한다

또한 dual encoder로 image-text retrieval을 위한 bidirectional loss 와 text-text retrieval을 위한 loss를 사용해 evaluation한다.

-> pre-trained BERT , pre-trained EfficientNet(B4)

-> small gain for image-text Task -> large gain for text-text task -> degradation for image-image task

model trades capacity for better text encoding