Abstract:Phytoplankton are a crucial component of aquatic ecosystems, and effective monitoring of them can provide valuable insights into ocean environments and ecosystem changes. Traditional phytoplankton monitoring methods are often complex and lack timely analysis. Therefore, deep learning algorithms offer a promising approach for automated phytoplankton monitoring. However, the lack of large-scale, high-quality training samples has become a major bottleneck in advancing phytoplankton tracking. In this paper, we propose a challenging benchmark dataset, Multiple Phytoplankton Tracking (MPT), which covers diverse background information and variations in motion during observation. The dataset includes 27 species of phytoplankton and zooplankton, 14 different backgrounds to simulate diverse and complex underwater environments, and a total of 140 videos. To enable accurate real-time observation of phytoplankton, we introduce a multi-object tracking method, Deviation-Corrected Multi-Scale Feature Fusion Tracker(DSFT), which addresses issues such as focus shifts during tracking and the loss of small target information when computing frame-to-frame similarity. Specifically, we introduce an additional feature extractor to predict the residuals of the standard feature extractor's output, and compute multi-scale frame-to-frame similarity based on features from different layers of the extractor. Extensive experiments on the MPT have demonstrated the validity of the dataset and the superiority of DSFT in tracking phytoplankton, providing an effective solution for phytoplankton monitoring.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are as follows: Current traditional methods for phytoplankton monitoring are complex and lack real - time analysis capabilities, and existing multi - target tracking algorithms are difficult to accurately track phytoplankton in complex underwater environments. Specifically, there are mainly two problems: 1. **Overlap and background interference**: When individual phytoplankton overlap with each other or with impurities, the attention of traditional multi - target tracking algorithms may be inappropriately shifted, resulting in tracking errors. 2. **Small - target information loss**: When calculating the similarity between consecutive frames, using only the deepest feature map for calculation will lead to the loss of small - target information, which is particularly disadvantageous for phytoplankton tracking because there are significant size differences between different species. To solve these problems, the paper proposes the following solutions: ### 1. Construct a large - scale multi - phytoplankton tracking benchmark dataset (MPT) - **Diversity**: The MPT dataset contains 27 species of phytoplankton and zooplankton, and 14 different background images to simulate diverse and complex underwater environments. - **Scale**: There are a total of 140 videos in the dataset, providing rich training samples. - **Quality**: All videos are of high resolution (4K), ensuring high - quality data collection. ### 2. Propose a new multi - target tracking framework (DSFT) #### 2.1 Deviation correction method (DCM) - **Problem**: The appearance of phytoplankton is often similar to the underwater environment. When they overlap or are partially overlapped, the attention of the algorithm will be inappropriately shifted. - **Solution**: Introduce an auxiliary feature extractor, predict the residual of the output of the main feature extractor, and add the residual to the main feature map, thereby correcting the deviation and ensuring that the algorithm focuses on a single target. #### 2.2 Multi - scale feature similarity fusion (MFSF) - **Problem**: Phytoplankton of different species have significant size differences. When the algorithm focuses on larger individuals, it is easy to ignore the features of smaller individuals. - **Solution**: During the feature extraction process, extract shallow - layer, middle - layer, and deep - layer features, corresponding to small, medium, and large objects respectively. Then calculate the similarity matrices of consecutive frames on each feature layer and fuse the results to prevent the loss of small - target information. ### Summary By constructing the MPT dataset and proposing the DSFT framework, this paper aims to solve the limitations of existing phytoplankton monitoring methods and provide a more efficient and accurate automatic phytoplankton monitoring scheme. The experimental results show the effectiveness of the MPT dataset and the superiority of the DSFT framework in the phytoplankton tracking task, providing strong support for the real - time monitoring of the marine ecosystem. \[ \text{MOTA} = \frac{\sum_{t = 1}^{T}(1-\frac{\text{FP}_t+\text{FN}_t+\text{IDs}_t}{\text{GT}_t})}{T} \] where: - \( \text{FP}_t \) is the number of false positives in the \( t \)-th frame - \( \text{FN}_t \) is the number of false negatives in the \( t \)-th frame - \( \text{IDs}_t \) is the number of identity switches in the \( t \)-th frame - \( \text{GT}_t \) is the number of real targets in the \( t \)-th frame - \( T \) is the total number of frames These improvements make phytoplankton monitoring more accurate and reliable.

MPT: A Large-scale Multi-Phytoplankton Tracking Benchmark

A Robust Underwater Multiclass Fish-School Tracking Algorithm

PhyTracker: An Online Tracker for Phytoplankton

Multi-class Fish Stock Statistics Technology Based on Object Classification and Tracking Algorithm

YoloXT: A Object Detection Algorithm for Marine Benthos

Underwater Target Tracking Based on the Feature-Aided GM-PHD Method

Deep-Learning-Based Automated Tracking and Counting of Living Plankton in Natural Aquatic Environments

WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark

FishTrack23: An Ensemble Underwater Dataset for Multi-Object Tracking

Improving Underwater Visual Tracking With a Large Scale Dataset and Image Enhancement

Multiple Dense Particle Tracking In Fluorescence Microscopy Images Based On Multidimensional Assignment

LMD-TShip⋆: Vision Based Large-Scale Maritime Ship Tracking Benchmark for Autonomous Navigation Applications

TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild

FMRFT: Fusion Mamba and DETR for Query Time Sequence Intersection Fish Tracking

Applying Feature-Similarity-metrics for Long-Tailed Problem of Phytoplankton Microscopic Images Classification

Multi-Target Deep Learning for Algal Detection and Classification

Underwater Object Tracker: UOSTrack for Marine Organism Grasping of Underwater Vehicles

Coastal Fisheries Resource Monitoring Through A Deep Learning-Based Underwater Video Analysis

DAPlankton: Benchmark Dataset for Multi-instrument Plankton Recognition via Fine-grained Domain Adaptation

The Caltech Fish Counting Dataset: A Benchmark for Multiple-Object Tracking and Counting

Underwater Long-Term Object Tracker for Marine Organism Capture