RT-DETR
RT-DETR at World Data Ocean is a file of 4 stories. The newest of them: “Counting shrimp with calibrated precision through enhanced acoustic detection”, “Maintaining Reliable Underwater Object Detection Amidst Challenging Conditions”, and “Real-Time Sonar Target Detection Enhanced with Innovative Edge-Frequency Transformer”. Counting shrimp with calibrated precision through enhanced acoustic detection is now measurable. Underwater vision systems rarely fail gracefully; they fail when clarity drops. A Global Hub for Ocean Intelligence 4 World Data Ocean is a centralized digital platform where researchers, scientists, and ocean enthusiasts converge to… The list below is every RT-DETR story on World Data Ocean, newest first.
Counting shrimp with calibrated precision through enhanced acoustic detection
Counting shrimp with calibrated precision through enhanced acoustic detection is now measurable. Small size, dense aggregation, and background clutter have long defeated conventional models. Our improved RT-DETR framework, integrating Efficient Multi-Scale Attention, a dedicated P2 detection layer, and combined Focal-WIoU loss, achieves a mean mAP@50 of 0.844. Paired with DeepSORT tracking, automated counting reaches 83.32% accuracy across video clips. This is a validated, empirical step toward real-world intake safety monitoring.
Maintaining Reliable Underwater Object Detection Amidst Challenging Conditions
Underwater vision systems rarely fail gracefully; they fail when clarity drops. Turbidity, color loss, and weak boundaries degrade the encoder features that RT-DETR-style detectors rely on to rank and initialize decoder queries. DQR-RTDETR confronts this directly, introducing a Candidate Reliability Prior to estimate trustworthiness from encoder memory, then recalibrating top-K scores and selected-query features with Stability-Preserving Query Reliability Recalibration. The results are measurable: mAP@0.5:0.95 rises from 0.6998 to 0.7444 on SeaClear, with only 0.05 million added parameters. This is a compact
Real-Time Sonar Target Detection Enhanced with Innovative Edge-Frequency Transformer
Side-scan sonar images arrive blurred and noise-heavy, a reality that complicates underwater target recognition. The proposed LEF-RT-DETR framework meets this directly: it sharpens edge perception through Gaussian-enhanced feature extraction while separating signal from noise in the frequency-spatial domain. The reported gains, 4.3% in AP and 5.3% in AP50 over RT-DETR, come alongside a meaningful reduction in parameters and computation. That combination matters for real-time operations.
Advancing Underwater Plastic Detection with Integrated Data and AI.
Underwater plastic debris is small, low-contrast, and easily lost in visually cluttered environments. That is why the improved RF-DETR detector developed here matters. By integrating frequency-aware feature reweighting with adaptive query strategies, the model sharpens weak-texture details and boosts localization precision. On the TrashCan dataset, it reached 93.11% precision, and its transfer to DeepTrash shows strong cross-dataset generalization. The results are practical, measurable, and directly relevant to monitoring efforts.