Large web crawl datasets have already played an important role in learning multimodal features with high generalization capabilities. However, there are still very limited studies investigating the details or improvements of data design. Recently, a DataComp challenge has been designed to propose the best training data with the fixed models. This paper presents our solution to both filtering track and BYOD track of the DataComp challenge. Our solution adopts large multimodal models CLIP and BLIP-2 to filter and modify web crawl data, and utilize external datasets along with a bag of tricks to improve the data quality. Experiments show our solution significantly outperforms DataComp baselines (filtering track: 6.6% improvement, BYOD track: 48.5% improvement).
In this study, we introduce a low cost method for generating descriptions from images containing novel objects. Generally, constructing a model, which can explain images with novel objects, is costly because of the following: (1) collecting a large amount of data for each category, and (2) retraining the entire system. If humans see a small number of novel objects, they are able to estimate their properties by associating their appearance with known objects. Accordingly, we propose a method that can explain images with novel objects without retraining using the word embeddings of the objects estimated from only a small number of image features of the objects. The method can be integrated with general image-captioning models. The experimental results show the effectiveness of our approach.
This paper addresses the generation of referring expressions that not only refer to objects correctly but also ease human comprehension. As the composition of an image becomes more complicated and a target becomes relatively less salient, identifying referred objects comes more difficult. However, the existing studies regarded all sentences that refer to objects correctly as equally good, ignoring whether they are easily understood by humans. If the target is not salient, humans utilize relationships with the salient contexts around it to help listeners to comprehend it better. To derive these information from human annotations, our model is designed to extract information from the inside and outside of the target. Moreover, we regard that sentences that are easily understood are those that are comprehended correctly and quickly by humans. We optimized it by using the time required to locate the referred objects by humans and their accuracies. To evaluate our system, we created a new referring expression dataset whose images were acquired from Grand Theft Auto V (GTA V), limiting targets to persons. Our proposed method outperformed previous methods both on machine evaluation and on crowd-sourced human evaluation. The source code and dataset will be available soon.