TI-UIED: Task-Inspired Underwater Enhancement Dataset

• TI-UIED is an underwater image enhancement dataset that selects reference images based on downstream task performance, such as semantic segmentation.
• It uses an automatic construction protocol with candidate enhancement methods, mIoU scoring, and majority voting to determine the optimal enhancement per raw image.
• Empirical validation shows improvements in segmentation metrics (Dice and mIoU) over perception-based datasets, underscoring its task-aligned design.

Searching arXiv for the cited TI-UIED source and closely related context papers. arXivSearch results for query "(Lin et al., 2 Mar 2026)":

• (Lin et al., 2 Mar 2026): "Downstream Task Inspired Underwater Image Enhancement: A Perception-Aware Study from Dataset Construction to Network Design" (published 2026-03-02) arXivSearch results for query "(Cong et al., 2024)":
• (Cong et al., 2024): "A Comprehensive Survey on Underwater Image Enhancement Based on Deep Learning" (published 2024-05-30) Task-Inspired UIE Dataset (TI-UIED) is an underwater image enhancement dataset introduced as a downstream-task inspired alternative to perception-centered UIE benchmarks. It is presented as a dataset in which the enhancement reference for each raw image is selected not by human visual preference, but by its utility for downstream recognition, especially semantic segmentation. In the source paper, TI-UIED is introduced alongside DTI-UIE (“Downstream Task-Inspired Underwater Image Enhancement”) and is defined through an automatic construction process based on candidate enhancement methods, multiple semantic segmentation networks, mIoU scoring, and majority voting (Lin et al., 2 Mar 2026).

1. Concept and motivation

TI-UIED was introduced to address a mismatch emphasized by the DTI-UIE paper: many existing underwater image enhancement methods and datasets are optimized for human-perceived image quality, while downstream recognition tasks depend on local structures, boundaries, textures, and other high-frequency cues that may not be preserved by visually pleasing enhancement targets (Lin et al., 2 Mar 2026).

The dataset is explicitly contrasted with prior UIE datasets such as UIEB, SUIM-E, and LSUI. Those datasets are described as perception-centered, typically relying on synthetic degradation or on selecting the preferred image from a pool of enhancement outputs through human majority voting. TI-UIED replaces that preference criterion with downstream-task performance. The paper’s central argument is that UIE supervision should be aligned with whether enhancement benefits recognition tasks such as semantic segmentation, object detection, and instance segmentation, rather than with subjective appearance alone (Lin et al., 2 Mar 2026).

This places TI-UIED within a broader shift already visible in the UIE literature. A 2024 survey states that “The UIE task can be used as an upstream task for other image understanding tasks,” and explicitly identifies object detection, feature matching, saliency detection, and semantic segmentation as downstream evaluation targets. The same survey also notes that enhancement can even suppress downstream performance in some settings, which is the gap TI-UIED is designed to target (Cong et al., 2024).

2. Automatic construction protocol

TI-UIED is built on the SUIM dataset and uses 1635 real underwater images with pixel-level semantic labels as the raw image source (Lin et al., 2 Mar 2026). The paper states that SUIM is chosen because it contains a wide variety of scenes, water types, lighting conditions, and target types.

For each raw image, the construction process generates enhancement candidates using 9 UIE methods, together with the raw image as an additional candidate. The nine enhancement methods are divided into two groups:

• Traditional methods: CLAHE, Fusion, ULAP, WWPE, WB
• Deep learning-based methods: SGUIE, TOPAL, TUDA, UColor

The paper explicitly states that the raw image is allowed to win the selection process. If the raw image yields the best downstream performance, it is used directly as the reference image. This means TI-UIED does not assume that enhancement is always beneficial (Lin et al., 2 Mar 2026).

The ranking mechanism is based on semantic segmentation. The source paper states: “we select the SUIM dataset as the target set, and make usage of the UIIS dataset as the training set to train various semantic segmentation networks.” The segmentation models are trained on enhanced UIIS images and tested on enhanced SUIM images. The seven segmentation architectures used in this selection stage are:

1. UNet
2. Segformer
3. PSPNet
4. DeepLabV3
5. DeepLabV3+
6. MaNet
7. FPN

To reduce randomness, each training process is repeated 3 times. The paper gives the total number of trained models as

$(5+4+1)\times 7 \times 3$

which equals 210 segmentation models (Lin et al., 2 Mar 2026).

The procedural selection rule is described as follows. First, multiple enhanced candidates are generated. Next, segmentation models trained on the corresponding enhancement style are evaluated on SUIM. Then, for each raw image, semantic segmentation performance is measured using mIoU. The enhancement result giving the strongest downstream performance is identified, and a majority voting mechanism across the task networks is used to determine the final selected reference. The exact tie-breaking or aggregation rule is not formalized in the paper beyond this description (Lin et al., 2 Mar 2026).

3. Dataset structure and inherited annotations

TI-UIED is effectively a paired UIE dataset. Each example consists of a raw underwater image and a selected enhancement reference image, with semantic annotations inherited from SUIM. The dataset therefore supports both enhancement supervision and downstream-task supervision.

The semantic annotation space consists of eight semantic categories inherited from SUIM:

1. water body
2. fish
3. reefs
4. aquatic plants
5. wrecks/ruins
6. human divers
7. robots
8. sea-floor

The paper’s comparison table summarizes TI-UIED with the following metadata: Year: 2025, Img. Num.: 1635, GT raw generation method: Majority-voting from Enc by downstream tasks, Enhance methods used: 5 Trad + 4 DL methods, Objective: Downstream-task inspired (Lin et al., 2 Mar 2026).

A concise summary is given below.

Item	TI-UIED specification
Raw image source	SUIM
Number of images	1635
Image type	Real underwater images
Reference selection basis	mIoU with majority voting
Candidate pool	5 traditional + 4 DL methods + raw image
Semantic labels	8 classes from SUIM

The split protocol is described relative to SUIM rather than as a separately enumerated TI-UIED partition. The paper states that, when SUIM is used to construct TI-UIED, only the training split of the original SUIM dataset is involved in dataset construction and enhancement model training, while evaluation is performed on the original test split of SUIM. Exact train/validation/test counts for TI-UIED are not specified (Lin et al., 2 Mar 2026).

The paper also specifies an input resolution used for model training: all images are resized to $256 \times 256$. It does not report the native image resolutions of the dataset itself (Lin et al., 2 Mar 2026).

4. Role in the DTI-UIE framework

TI-UIED is not only a benchmark; it is the supervision source used to train DTI-UIE. In the source formulation, the training tuple contains a raw image $I_{raw}$, an enhancement target $I_{gt}$, and a semantic segmentation mask $M_{gt}$ (Lin et al., 2 Mar 2026).

The paper defines the coupled enhancement-and-task pipeline as:

$$\hat{T}_{pri} =\mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$

$$\hat{I}_{enc} =\mathbf{Enc}(I_{raw},\hat{T}_{pri};\theta_{enc}),$$

$$\hat{M}_{enc} =\mathbf{Seg}_{task}(\hat{I}_{enc};\theta_{task}).$$

The corresponding training objective is written as

$\underset{\theta_{enc},\theta_{pri},\theta_{task}{\min} \ \mathcal{L}_{enc}(\hat{I}_{enc},I_{gt})+\mathcal{L}_{task}(\hat{M}_{enc},M_{gt}).$

This tight coupling explains the dataset’s design. TI-UIED provides references already selected to preserve recognition-relevant content, and DTI-UIE reinforces that alignment through architecture and loss design. The enhancement network is a dual-branch model consisting of a Feature Restoration Branch (FRB) and a Detail Enhancement Branch (DEB), while task priors are injected through the Task-Aware Conv-attention Transformer Block (TA-CTB) (Lin et al., 2 Mar 2026).

The paper also introduces a Task-Driven Perceptual loss:

$$\mathcal{L}_{TDP}=\left \| \mathbf{Seg}_{task}^{fe}(\hat{I}_{enc})-\mathbf{Seg}_{task}^{fe}(I_{gt})\right \| _1,$$

where the feature extractor belongs to the downstream task network. This means the predicted enhancement is compared to the TI-UIED reference not only at the pixel level but also in a feature space optimized for the downstream task (Lin et al., 2 Mar 2026).

A plausible implication is that TI-UIED is best understood as a data-level intervention, while DTI-UIE adds a model-level intervention. The paper explicitly treats them as complementary: the dataset aligns the target image with downstream utility, and the network plus losses align the learned representation with the same objective.

5. Empirical validation

The most direct empirical evidence for TI-UIED is an ablation comparing the same enhancement architecture trained on SUIM-E versus trained on TI-UIED. Table 10 in the paper reports:

• DTI-UIE trained on SUIM-E: Dice 70.47, mIoU 65.18
• DTI-UIE trained on TI-UIED: Dice 74.48, mIoU 67.61

This corresponds to +4.01 Dice and +2.43 mIoU when TI-UIED is used instead of the human-oriented SUIM-E reference construction (Lin et al., 2 Mar 2026). Within the paper’s argument, this is the clearest dataset-specific validation that task-inspired reference selection changes the supervision in a useful way.

The paper also evaluates whether TI-UIED helps other UIE models. It reports:

• WfDiff source pretrained: Dice 72.37, mIoU 66.09
• WfDiff trained on TI-UIED: Dice 73.61, mIoU 66.25
• CCMSR source pretrained: Dice 72.58, mIoU 65.47
• CCMSR trained on TI-UIED: Dice 70.50, mIoU 63.96

This suggests that TI-UIED can improve some existing methods, but that dataset supervision alone does not guarantee gains for architectures designed primarily around human visual quality (Lin et al., 2 Mar 2026).

Beyond the dataset-specific ablations, the paper reports downstream improvements for the full DTI-UIE system on three task families:

• Semantic segmentation on SUIM and LIACi
• Object detection on RUOD
• Instance segmentation on UIIS

These experiments validate the broader task-inspired design, although they evaluate the combined dataset-and-model pipeline rather than TI-UIED in isolation. An important detail is that TI-UIED is constructed using semantic segmentation only, yet the resulting enhancement transfers to object detection and instance segmentation as well (Lin et al., 2 Mar 2026).

The paper also reports that DTI-UIE is only comparable, not best, on perception-oriented no-reference metrics such as UIQM, UCIQE, and NUIQ. This is used to support the claim that conventional image-quality metrics reflect human perception more strongly than downstream task utility. This suggests that TI-UIED deliberately encodes a different target notion of “good” enhancement than earlier perception-centered datasets (Lin et al., 2 Mar 2026).

6. Position in the literature and limitations

Within the UIE literature, TI-UIED occupies a distinct position. The 2024 survey on underwater image enhancement does not explicitly mention TI-UIED, but it provides the taxonomic background against which TI-UIED is legible: most benchmark datasets are organized around paired versus unpaired construction, real versus synthetic data, and perceptual evaluation, while task-oriented evaluation remains comparatively underdeveloped (Cong et al., 2024). TI-UIED can therefore be read as a dataset-level answer to a gap that the survey already identified.

The paper also acknowledges a central limitation: TI-UIED is segmentation-driven. Although multiple segmentation architectures are used in construction, the reference selection process is still tied to one downstream task family. The authors argue that segmentation is a dense prediction task requiring both global semantics and local structural fidelity, but they explicitly note that other downstream tasks could rank candidates differently (Lin et al., 2 Mar 2026).

This implies a specific form of dataset bias. The selected references may favor image characteristics that are optimal for segmentation-style recognition, even when those characteristics are not universally optimal for other underwater tasks. The paper therefore proposes future extensions that would incorporate multiple downstream tasks more explicitly in dataset construction and learning, including joint use of segmentation, detection, and instance segmentation (Lin et al., 2 Mar 2026).

Other limitations are procedural rather than conceptual. The paper does not specify an exact vote-counting rule, a tie-breaking policy, or an explicit ranking formula beyond mIoU and majority voting. It also does not provide exact train/validation/test counts for TI-UIED as a standalone release artifact, and it does not state a separate dataset download URL in the provided text, although it announces public code release for the accompanying DTI-UIE project (Lin et al., 2 Mar 2026).

Taken together, TI-UIED represents a redefinition of UIE supervision: not enhancement toward human preference, but enhancement toward downstream recognizability. Its main significance lies in converting that principle into a concrete paired dataset built from real underwater images, task-network-based selection, and inherited dense semantic labels, thereby linking dataset construction directly to the operational requirements of underwater recognition systems (Lin et al., 2 Mar 2026).