• Journal of Information Processing Systems
• /
• v.16 no.6
• /
• pp.1250-1260
• /
• 2020

This study proposes a novel deep neural network model that can accurately detect objects and their relationships in an image and represent them as a scene graph. The proposed model utilizes several multimodal features, including linguistic features and visual context features, to accurately detect objects and relationships. In addition, in the proposed model, context features are embedded using graph neural networks to depict the dependencies between two related objects in the context feature vector. This study demonstrates the effectiveness of the proposed model through comparative experiments using the Visual Genome benchmark dataset.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.10
• /
• pp.2643-2657
• /
• 2023

Scene graphs are structured representations that can clearly convey objects and the relationships between them, but are often heavily biased due to the highly skewed, long-tailed relational labeling in the dataset. Indeed, the visual world itself and its descriptions are biased. Therefore, Unbiased Scene Graph Generation (USGG) prefers to train models to eliminate long-tail effects as much as possible, rather than altering the dataset directly. To this end, we propose Geometric and Semantic Improvement (GSI) for USGG to mitigate this issue. First, to fully exploit the feature information in the images, geometric dimension and semantic dimension enhancement modules are designed. The geometric module is designed from the perspective that the position information between neighboring object pairs will affect each other, which can improve the recall rate of the overall relationship in the dataset. The semantic module further processes the embedded word vector, which can enhance the acquisition of semantic information. Then, to improve the recall rate of the tail data, the Class Balanced Seesaw Loss (CBSLoss) is designed for the tail data. The recall rate of the prediction is improved by penalizing the body or tail relations that are judged incorrectly in the dataset. The experimental findings demonstrate that the GSI method performs better than mainstream models in terms of the mean Recall@K (mR@K) metric in three tasks. The long-tailed imbalance in the Visual Genome 150 (VG150) dataset is addressed better using the GSI method than by most of the existing methods.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.12
• /
• pp.3383-3397
• /
• 2023

Scene graphs serve as semantic abstractions of images and play a crucial role in enhancing visual comprehension and reasoning. However, the performance of Scene Graph Generation is often compromised when working with biased data in real-world situations. While many existing systems focus on a single stage of learning for both feature extraction and classification, some employ Class-Balancing strategies, such as Re-weighting, Data Resampling, and Transfer Learning from head to tail. In this paper, we propose a novel approach that decouples the feature extraction and classification phases of the scene graph generation process. For feature extraction, we leverage a transformer-based architecture and design an adaptive calibration function specifically for predicate classification. This function enables us to dynamically adjust the classification scores for each predicate category. Additionally, we introduce a Distribution Alignment technique that effectively balances the class distribution after the feature extraction phase reaches a stable state, thereby facilitating the retraining of the classification head. Importantly, our Distribution Alignment strategy is model-independent and does not require additional supervision, making it applicable to a wide range of SGG models. Using the scene graph diagnostic toolkit on Visual Genome and several popular models, we achieved significant improvements over the previous state-of-the-art methods with our model. Compared to the TDE model, our model improved mR@100 by 70.5% for PredCls, by 84.0% for SGCls, and by 97.6% for SGDet tasks.