Less Is More: Linear Layers on CLIP Features as Powerful VizWiz Model
This work addresses computational efficiency and training difficulty for researchers and practitioners in visual question answering, but it is incremental as it builds on existing CLIP features with simple modifications.
The paper tackled the problem of high complexity in multi-modality architectures for visual question answering by proposing a CLIP-based model that avoids fine-tuning feature extractors and uses a linear classifier on concatenated features, achieving 60.15% accuracy on answer prediction and 83.78% AP on answerability prediction in the VizWiz 2022 challenge.
Current architectures for multi-modality tasks such as visual question answering suffer from their high complexity. As a result, these architectures are difficult to train and require high computational resources. To address these problems we present a CLIP-based architecture that does not require any fine-tuning of the feature extractors. A simple linear classifier is used on the concatenated features of the image and text encoder. During training an auxiliary loss is added which operates on the answer types. The resulting classification is then used as an attention gate on the answer class selection. On the VizWiz 2022 Visual Question Answering Challenge we achieve 60.15 % accuracy on Task 1: Predict Answer to a Visual Question and AP score of 83.78 % on Task 2: Predict Answerability of a Visual Question.