Fine Tuning Panoptic Scene Graph Generation

The understanding of scene graph generation evolves from bounding box approaches to segmentation techniques. The initial development in scene graph generation benchmarks aims to enhance the understanding of contextual information in visual images. Subsequent advancements using panoptic scene graph generation improve the precision of contextual information in visual scenes, facilitated by segmentation that minimizes information loss through masking processes. To comprehend and advance understanding of visual scenes, it is crucial to grasp the rationale and architecture of panoptic scene graph generation. Therefore, this study analyzes the panoptic scene graph generation process by fine-tuning hyperparameters, specifically the learning rate using cosine annealing and Nesterov momentum. Additionally, the integration of cosine annealing and Nesterov momentum shows effectiveness in accelerating the process and reducing the loss function, resulting in higher mean Recall (mR). In this study, the highest mR value was observed at epoch 7, with a training loss of 7.37, test loss of 18.23, and mR of 17.40.