The field of computer vision is moving towards more robust and efficient methods for human-object interaction detection and image segmentation. Recent developments have focused on improving the resilience of models to real-world challenges such as environmental variability, occlusion, and noise. Researchers are also exploring new approaches to prompt engineering, dynamic attention mechanisms, and semi-supervised learning to enhance model performance. Notable advancements include the use of wavelet attention-like backbones, ray-based encoder architectures, and novel loss functions for bounding box regression. These innovations have the potential to significantly improve the accuracy and efficiency of computer vision systems in various applications, including robot-human assistance, medical imaging, and remote sensing. Noteworthy papers include: RoHOI, which introduces a robustness benchmark for human-object interaction detection and proposes a semantic-aware masking-based progressive learning strategy. Inter2Former, which presents a dynamic hybrid attention mechanism for efficient high-precision interactive segmentation. DEARLi, which devises a novel semi-supervised panoptic approach fueled by two dedicated foundation models to enhance recognition and localization.