Reproduce “HFP-SAM: Hierarchical Frequency Prompted SAM for Efficient Marine Animal Segmentation” on OpenResearch

About this paper

Pingping Zhang, Tianyu Yan, Yuhao Wang, Yang Liu, Tongdan Tang, Yili Ma, Long Lv, Feng Tian and 2 more

HFP-SAM introduces a novel framework for high-performance marine animal segmentation by integrating hierarchical frequency domain prompting and a Full-View Mamba decoder with the Segment Anything Model. The approach consistently outperforms existing methods by 4-10% across four challenging marine datasets and demonstrates strong generalization capabilities to other diverse segmentation tasks.

Key insights

• Leveraging frequency domain information through adapters and intelligent prompt selection is highly effective for adapting general-purpose foundation models like SAM to specialized, noisy environments such as marine animal segmentation.
• Combining frequency-aware point prompts with a coarse segmentation mask provides complementary boundary-level and shape-level cues, resulting in more accurate and robust segmentation without requiring complex prompt generation networks.
• The Mamba architecture, when integrated into a decoder, efficiently models long-range spatial and channel dependencies, significantly improving detail preservation and effective receptive field compared to simpler decoders.

Summary by alphaXiv.