Answers are grounded in this paper and cite their sources.
CrabFormer demonstrates enhanced crab identification and pose estimation in automated seafood processing, suggesting improved efficiency.
• CrabFormer integrates RGB-D inputs for instance segmentation and pose estimation. • Dual-patch Swin-T backbone enhances RGB-D feature extraction. • CrabFormer outperforms state-of-the-art models, particularly in piled crab cases. • The model achieves competitive inference times with multitasking capabilities. • CrabFormer enhances automation potential for seafood front-end loading. CrabFormer is a multitasking transformer model developed to tackle the challenges of front-end loading in the automated processing of Chesapeake blue crabs. Existing methods often struggle to accurately identify crabs in chaotic, piled configurations, where occlusions and overlapping are common. CrabFormer addresses this by combining instance segmentation and keypoint prediction using RGB-D inputs to detect crabs and estimate their orientation. Utilizing a dual-patch Swin-T backbone, the model processes RGB and depth data separately, effectively capturing visual and geometric features. CrabFormer is evaluated on a custom dataset comprising discrete, overlapping, and piled crabs. It achieves a segmentation Average Precision (AP) of 69.76 and Average Recall (AR) of 78.54, while its keypoint prediction AP and AR are 62.01 and 85.48, respectively. The model outperforms state-of-the-art transformer-based segmentation and keypoint prediction models, particularly in the most complex piled cases, demonstrating improvements in both AP and AR. It excels in identifying the topmost crab in a pile, a key task for automated processing. Additionally, CrabFormer achieves competitive inference times while maintaining superior multitasking performance. These results highlight CrabFormer’s potential to enhance the automation of front-end loading in the seafood industry, reducing labor reliance and improving operational efficiency. Future work will expand the dataset and explore the model’s applicability to other crustaceans with similar morphological complexities.
No takes yet. Share an insight, caveat, or question.
Ali et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: