Siqi Yin

Shukang Yin, Chaoyou Fu, Sirui Zhao et al.

Recent years have seen the success of Multimodal Large Language Models (MLLMs) in the domain of vision understanding. The success of these models can largely be attributed to the dominant scaling law, which states that larger parameter sizes and data volumes contribute to better performance. Notably, data scaling has been primarily driven by automatic data pipelines, which focus on the self-instruction of LLMs. The paradigm has been taken for granted for quite some time, but the study of the effectiveness of scaling with these data has been neglected for a long time. In this context, this work revisits scaling with synthetic data and focuses on developing video-LLMs from a data-centric perspective. Our primary study approach involves fine-tuning pre-trained image-LLMs with video data and examining learning efficiency through data scaling. Results from our preliminary experiments reveal a low learning efficiency phenomenon when simply scaling up video data samples, which, through our probing, can be ascribed to a lack of instruction diversity. Aiming at this issue, we propose a data augmentation method called Sparrow, which synthesizes video-like samples from pure text instruction data. Mixing these synthetic samples with the video data enables a more efficient training scheme. Through comprehensive experiments, we demonstrate that our proposed method achieves performance comparable to or even superior to that of baselines trained with significantly more samples. Meanwhile, we find that incorporating these synthetic samples can enhance the performance of long video understanding without requiring training on long video data. The code and data examples are available at https://github.com/VITA-MLLM/Sparrow.

Siqi Yin, Shaolei Liu, Manning Wang

Domain adaptation addresses the challenge of model performance degradation caused by domain gaps. In the typical setup for unsupervised domain adaptation, labeled data from a source domain and unlabeled data from a target domain are used to train a target model. However, access to labeled source domain data, particularly in medical datasets, can be restricted due to privacy policies. As a result, research has increasingly shifted to source-free domain adaptation (SFDA), which requires only a pretrained model from the source domain and unlabeled data from the target domain data for adaptation. Existing SFDA methods often rely on domain-specific image style translation and self-supervision techniques to bridge the domain gap and train the target domain model. However, the quality of domain-specific style-translated images and pseudo-labels produced by these methods still leaves room for improvement. Moreover, training the entire model during adaptation can be inefficient under limited supervision. In this paper, we propose a novel SFDA framework to address these challenges. Specifically, to effectively mitigate the impact of domain gap in the initial training phase, we introduce preadaptation to generate a preadapted model, which serves as an initialization of target model and allows for the generation of high-quality enhanced pseudo-labels without introducing extra parameters. Additionally, we propose a data-dependent frequency prompt to more effectively translate target domain images into a source-like style. To further enhance adaptation, we employ a style-related layer fine-tuning strategy, specifically designed for SFDA, to train the target model using the prompted target domain images and pseudo-labels. Extensive experiments on cross-modality abdominal and cardiac SFDA segmentation tasks demonstrate that our proposed method outperforms existing state-of-the-art methods.