UW-DualDet: Underwater Object Detection Dataset for Illumination and Occlusion Scenes

A research paper proposes the Underwater Dual-modal Detection Network (UW-DualDet) for object detection in challenging underwater environments. The method, authored by Chao Zhang, incorporates RGB and depth data to address degradation from light scattering and dense occlusion. The paper reports a mean Average Precision (mAP) of 86.7% and an inference speed of 116 FPS, with results from clear, turbid, and low-light deep water scenarios.

Use Cases

• Training object detection models for underwater robotics based on the described dual-modal (RGB-D) approach.
• Benchmarking detection algorithms in turbid or low-light water conditions based on the reported AP scores.
• Developing real-time underwater perception systems based on the 116 FPS inference speed mentioned.
• Studying feature fusion techniques for degraded images based on the described Underwater Denoise Feature Fusion (UDFF) module.

Strengths

• The proposed model achieved a reported mean Average Precision (mAP) of 86.7%.
• Model inference speed is reported as 116 FPS, meeting real-time requirements.
• Performance is quantified across three distinct underwater scenarios: clear water (88.2% AP₅₀), turbid water (79.6% AP₅₀), and low-light deep water (75.3% AP₅₀).

Limitations

• The dataset itself is not described; the 469.4 KB file is a PDF research paper, not the underlying image or depth data.
• Column-level documentation is absent; field semantics must be inferred after download if data is present.
• Row count is unknown, which may limit suitability assessment for model training.

Provenance

Source

Chao Zhang via figshare.

Collection Method

Methodology described in the paper involves a dual-modal network (UW-DualDet) using RGB and depth information.

Freshness

Last updated 2026-06-03 06:01:17; freshness should be verified.