MFT25: Underwater Fish Tracking Benchmark 2025

• MFT25 is an underwater multiple fish tracking benchmark featuring 15 video sequences, 48,066 frames, and 223 trajectories for advanced ecological and aquaculture research.
• The dataset uses genuine footage from diverse aquaculture environments with precise manual annotations to enhance studies of fish behavior and motion dynamics.
• The SU-T tracker integrates a specialized Unscented Kalman Filter and FishIoU metric to improve association accuracy over conventional terrestrial MOT approaches.

Multiple Fish Tracking Dataset 2025 (MFT25) is an underwater multiple object tracking benchmark introduced together with the Scale-aware and Unscented Tracker (SU-T) in 2025. It is presented as the first comprehensive dataset specifically designed for underwater multiple fish tracking, with 15 diverse video sequences, 48,066 frames, 408,578 manually annotated bounding boxes, and 223 distinct fish trajectories. The benchmark targets conditions that are central to marine ecology and aquaculture but poorly served by terrestrial MOT resources, including occlusions, similar appearances, erratic motion patterns, visibility fluctuations, turbidity, and lighting changes (Li et al., 8 Jul 2025).

1. Definition and research motivation

MFT25 was created to fill what its authors describe as a major gap in the MOT ecosystem: established benchmarks primarily focus on people, vehicles, or other land-based targets, whereas fish tracking datasets had been too small, too low-quality, or not standardized enough for modern tracking research (Li et al., 8 Jul 2025). The benchmark is explicitly framed around underwater multiple fish tracking rather than generic underwater MOT, with the underlying claim that fish motion and appearance differ sufficiently from terrestrial targets to require a dedicated dataset and dedicated tracking logic.

The stated application space is broad. The benchmark is motivated by marine ecology, aquaculture optimization, fishery resource management, endangered species protection, and ecological modeling. In this framing, identity-preserving fish tracking is not merely a vision benchmark problem; it is an enabling layer for behavior monitoring, schooling analysis, and long-term non-invasive observation.

A central conceptual point in the paper is that fish tracking is fundamentally different from terrestrial tracking in both appearance and motion. Fish often exhibit high morphological similarity, especially within the same species or developmental stage, while underwater imaging introduces reflections, blur, turbidity, and lighting instability. At the same time, fish trajectories are highly non-linear, with frequent and abrupt direction changes, rapid accelerations, and dense group interactions. This combination motivates both the benchmark design and the specialized SU-T baseline (Li et al., 8 Jul 2025).

2. Dataset composition and acquisition conditions

MFT25 was collected with Canon EOS R6 and Sony α7M3 cameras in diverse aquaculture environments, specifically industrial circulating water aquaculture ponds and controlled laboratory tanks. The recordings include both overhead and horizontal viewpoints and span illumination conditions from daylight to nocturnal settings. The paper emphasizes that the dataset uses real, unaltered footage rather than synthetic augmentation, preserving natural underwater complexity (Li et al., 8 Jul 2025).

The dataset includes multiple fish species with different morphologies, including commercially valuable groupers and ornamental koi at various developmental stages. The use of species and developmental diversity is intended to increase appearance diversity while also making identity association harder.

Attribute	Value
Video sequences	15
Frames	48,066
Manually annotated bounding boxes	408,578
Distinct fish trajectories	223
Cameras	Canon EOS R6; Sony α7M3
Environments	Industrial circulating water aquaculture ponds; controlled laboratory tanks
Viewpoints	Overhead; horizontal
Illumination	Daylight to nocturnal settings

The paper explicitly notes that the scale of MFT25 represents roughly 2.6–9.1× more annotated instances than earlier fish benchmarks. It also contrasts the dataset with terrestrial MOT datasets such as MOT17 and MOT20: whereas those datasets are broader in object category and scenario coverage, MFT25 is deliberately domain-specific and organized around fish behavior, underwater visibility issues, and fish morphology (Li et al., 8 Jul 2025).

3. Task characteristics and domain-specific difficulty

The benchmark’s difficulty is described along two axes: appearance ambiguity and motion irregularity. On the appearance side, many fish are visually similar to one another, especially within the same species or developmental stage. Underwater reflections, turbidity, blur, and illumination instability further weaken identity cues. On the motion side, fish do not follow the near-linear motion patterns common in pedestrian tracking; instead, they display chaotic directional changes, abrupt accelerations, and dense interactions (Li et al., 8 Jul 2025).

The paper supports this distinction with a statistical comparison between MFT25 and datasets for pedestrians, dancers, bees, and cells. It reports that fish in MFT25 show much higher angular velocity variation and higher average speed variability than terrestrial MOT datasets. MOT17 is used as a representative contrast case: pedestrian trajectories are described as relatively stable and near-linear, whereas MFT25 exhibits non-linear swimming behavior and chaotic directional changes.

These observations are used to argue that strong terrestrial trackers do not transfer directly. A standard motion prior that is adequate for pedestrians or vehicles becomes misaligned with fish kinematics, while appearance models inherit additional failure modes from underwater imaging. A plausible implication is that association errors on MFT25 arise not only from detector noise but also from a mismatch between terrestrial inductive biases and aquatic motion statistics.

4. Benchmark protocol and reported performance

For the benchmark comparison, all methods were evaluated on MFT25 using the same YOLOX-X detector and, for state-based detection-and-embedding models, the same FastReID-based Re-ID network, with consistent training configurations. Training was performed on a Tesla A100 GPU. The reported evaluation measures are HOTA, IDF1, MOTA, AssA, DetA, IDs, IDFP, IDFN, and Frag (Li et al., 8 Jul 2025).

The paper compares SU-T against FairMOT, CMFTNet, Deep-OC-SORT, OC-SORT, TFMFT, BoT-SORT, TransCenter, SORT, TrackFormer, TransTrack, ByteTrack, and HybridSORT. Its main empirical claim is that SU-T achieves the best overall performance on MFT25, especially in association-related metrics.

Tracker variant	HOTA	IDF1
SU-T	33.351	41.717
SU-T$^\dagger$ with Re-ID	34.067	44.643

The dagger variant is the strongest model reported in the paper. It also achieves 23.594 AssA and the lowest false negatives among the compared methods. The abstract rounds these values to 34.1 HOTA and 44.6 IDF1. The reported pattern is important: the gains are concentrated in identity continuity and association quality, which is consistent with the benchmark’s emphasis on visually similar fish and non-linear motion (Li et al., 8 Jul 2025).

The paper also reports that SU-T generalizes reasonably to terrestrial benchmarks, with 60.4 HOTA on MOT17 and 56.5 HOTA on MOT20. These cross-domain results are secondary to the MFT25 benchmark itself, but they are used to argue that fish-specific adaptations do not simply overfit aquatic footage.

5. SU-T framework and fish-specific association logic

SU-T is described as a lightweight SDE-style tracker with three main components: a detector, an association module, and an optional Re-ID module. The detector produces bounding boxes and confidence scores. The association module combines a custom similarity metric, FishIoU, with motion predictions from an Unscented Kalman Filter (UKF). The tracker then uses the Hungarian algorithm for assignment (Li et al., 8 Jul 2025).

The use of UKF is motivated by non-linear fish motion. The paper argues that a standard Kalman Filter is not ideal because fish trajectories include sudden directional changes and other non-linear swimming patterns. UKF is preferred because it propagates deterministic sigma points through non-linear transition and measurement models, thereby better accommodating abrupt motion variability.

FishIoU is the second major specialization. Standard IoU and box-overlap variants are treated as insufficient because fish are elongated, vary in scale, and can rotate or swim erratically in ways that make center distance alone unreliable. FishIoU combines standard IoU, central IoU, aspect ratio consistency, area ratio consistency, and a scale-aware center-distance penalty:

$$\text{FishIoU} = \omega_1 \cdot \text{IoU} + \omega_2 \cdot \text{cIoU} + \omega_3 \cdot \alpha_r + \omega_4 \cdot \alpha_a - \omega_5 \cdot s \cdot d_c$$

The paper gives default values $\alpha = 0.15$, $\beta = 0.3$, and $\gamma = 0.25$ for the central-region construction, and weights $\omega_1 = 1.0$, $\omega_2 = 0.3$, $\omega_3 = 0.1$, $\omega_4 = 0.2$, and $\omega_5 = 0.4$ for the final FishIoU score. The association procedure is multi-stage: high-confidence detections are matched first, low-confidence detections are used in a second round to recover occluded targets, and a last-chance association reconnects tracks using historical observations. The paper states that this design is inspired by ByteTrack and OC-SORT but adapted to fish-specific conditions (Li et al., 8 Jul 2025).

The ablation results support each specialized component. UKF performs better than standard KF, AKF, and STF for fish motion; FishIoU outperforms center distance, IoU, CIoU, DIoU, HMIoU, and GIoU; and the best Re-ID backbone is SBS-S101. These results reinforce the paper’s central argument that underwater fish tracking requires motion modeling and association criteria that are not inherited unchanged from terrestrial MOT.

6. Relation to earlier fish-tracking datasets and adjacent work

MFT25 is best understood within a developing lineage of fish-tracking resources rather than as an isolated benchmark. The paper explicitly contrasts it with Fish4Knowledge, DeepFish, SeaCLEF, BrackishMOT, 3D-ZeF, and MFT22, describing older datasets as limited by low resolution, poor visibility, inconsistent annotation formats, or restricted environmental diversity (Li et al., 8 Jul 2025).

A recurrent point of confusion concerns the relationship between MFT25 and earlier underwater MOT datasets. BrackishMOT, introduced in 2023, is a public underwater multi-object tracking benchmark focused on brackish-water scenes, especially tracking schools of small fish in turbid waters in Denmark. It contains 98 sequences captured in the wild, uses MOTChallenge-style annotations, and was presented at the time as the first and only dataset of its kind (Pedersen et al., 2023). However, the BrackishMOT paper does not mention “Multiple Fish Tracking Dataset 2025 (MFT25)” by name, nor does it indicate a 2025 release. The safest interpretation is that BrackishMOT is a related predecessor benchmark rather than the same resource.

A second distinct line is represented by the sturgeon-tracking dataset introduced with FMRFT. That work establishes a complex multi-scenario sturgeon tracking dataset for factory farming environments and reports 8,000 high-quality sturgeon fish tracking images in the abstract, while the experiments section states that 11,000 labeled video frames were obtained after frame splitting and automatic labeling. It is explicitly a sturgeon dataset paper rather than a paper about MFT25, and there is no mention of the MFT25 name or a 2025 benchmark release in that text (Yao et al., 2024).

A third adjacent contribution is the 2025 multi-view fish-tracking thesis that uses 182 videos, including 91 stereo videos, 84,464 frames, 3,796 tracks, ground-truth MOT .txt files, and stereo camera calibration files. That work is not a dataset paper for MFT25, but it is directly relevant as a methodological precursor for stereo matching, triangulation, and 3D fish-track reconstruction (Elchik et al., 22 May 2025).

Taken together, these neighboring works indicate a progression from species- or environment-specific fish MOT resources toward broader benchmark standardization. This suggests that MFT25’s distinctive role is not merely that it is newer, but that it combines larger scale, domain-specific evaluation, and a fish-specialized baseline in a single benchmark package.