Task-Aware Preprocessing Selection for Underwater Sparse 3D Reconstruction via Lightweight Machine Learning Under Grouped Evaluation Protocol

Abstract

Underwater image enhancement has been widely studied to improve visual quality; however, its impact on downstream geometric tasks such as sparse 3D reconstruction remains insufficiently understood. In particular, visually enhanced images do not necessarily lead to improved feature matching or reconstruction performance. This work addresses the problem of selecting appropriate preprocessing strategies for underwater Structure-from-Motion (SfM) pipelines from a task-oriented perspective. We propose a lightweight machine-learning-based preprocessing selector that predicts reconstruction performance from image statistics and recommends suitable enhancement strategies for each input sequence. To ensure reliable evaluation, we introduce a grouped leave-one-parent-sequence-out protocol that avoids overlap-induced bias common in clip-wise splitting. Experiments are conducted on challenging underwater datasets derived from the Real-world Underwater Image Enhancement (RUIE) benchmark, with the primary comparison variable defined as the number of reconstructed sparse 3D points. Supporting geometric variables, including the number of registered images, mean track length, and mean reprojection error, are recorded for interpretation. Results show that preprocessing choices significantly affect reconstruction outcomes and that the optimal strategy is scene-dependent. The proposed selector consistently improved over raw input on the evaluated grouped subset and remained competitive with a strong fixed preprocessing baseline. The grouped leave-one-parent-sequence-out protocol is intended to reduce overlap-induced bias common in clip-wise splitting and to provide a more conservative estimate of generalization. This work highlights the importance of task-aware preprocessing and reliable evaluation in underwater vision systems, offering practical insights for deploying enhancement strategies in real-world 3D reconstruction pipelines.

1. Introduction

Image-based 3D reconstruction is increasingly used in underwater archaeology, habitat monitoring, infrastructure inspection, and environmental documentation [1,2,3]. However, underwater imagery is affected by wavelength-dependent attenuation, backscatter, reduced contrast, blur, and color shift, which degrade both visual quality and geometric reliability [4,5,6,7,8,9]. These degradations are especially problematic for Structure-from-Motion (SfM) pipelines, whose performance depends on stable feature detection, matching, triangulation, and verification [1,2,3,10,11,12].

Underwater image enhancement has therefore developed rapidly, supported by benchmark datasets such as the Real-world Underwater Image Enhancement (RUIE) benchmark and the Underwater Image Enhancement Benchmark (UIEB) and by a broad range of restoration and learning-based enhancement methods [13,14,15,16,17,18,19]. At the same time, underwater image quality assessment has been increasingly standardized through no-reference metrics such as the Underwater Color Image Quality Evaluation (UCIQE) and the Underwater Image Quality Measure (UIQM), frequency-domain quality measures, and broader underwater image quality assessment (UIQA) frameworks [20,21,22,23]. Nevertheless, most enhancement studies still emphasize perceptual quality, whereas underwater robotics and photogrammetry ultimately require downstream geometric utility. Recent work has shown that radiometric preprocessing can alter underwater 3D reconstruction outcomes and that visually improved images do not always produce better feature matching or reconstruction [1,2,3,10,11,24].

1.1. Related Work

Relevant studies can be grouped into four lines. The first line concerns underwater image formation and physically grounded restoration, including polarization-based recovery, wavelength compensation, attenuation modeling, and water removal [4,5,6,7,8,9].

The second line focuses on underwater enhancement benchmarks and algorithms. RUIE and UIEB enabled more reproducible evaluation of enhancement methods under real underwater degradations, while subsequent methods explored prior-guided convolutional models, adversarial learning, color-space fusion, and multi-feature enhancement strategies [13,14,15,16,17,18,19].

The third line addresses underwater image quality assessment. UCIQE and UIQM remain widely used, while frequency-domain and more comprehensive no-reference assessment frameworks have been proposed to better characterize underwater image quality [20,21,22,23].

The fourth line concerns downstream task-oriented evaluation. Earlier studies showed that underwater radiometric preprocessing can change reconstruction quality [10,11], and more recent work has further highlighted the mismatch between perceptual improvement and geometric utility in feature matching, SLAM, and related vision tasks [1,2,3,24]. These findings motivate preprocessing selection as a task-aware decision problem rather than a universally fixed enhancement rule.

1.2. Study Objective

Against this background, this study addresses a practical question: when implementation complexity and computational cost must remain low, should an underwater subsequence be reconstructed directly from raw imagery or after lightweight preprocessing? To answer this question, four practical candidates are compared within a COLMAP-based sparse reconstruction pipeline: raw input, CLAHE, gray-world white balancing, and white-balance-plus-CLAHE [11,12,25].

To address these challenges, this paper makes the following contributions:

1. Task-aware preprocessing formulation for underwater 3D reconstruction. We reformulate underwater image enhancement as a task-driven preprocessing selection problem, explicitly linking image restoration strategies to downstream sparse 3D reconstruction performance rather than perceptual quality.

2. Lightweight machine-learning-based preprocessing selector. We develop a low-complexity regression-based selector that predicts reconstruction outcomes from image statistics and adaptively recommends preprocessing strategies, enabling practical deployment without heavy computational overhead.

3. Grouped evaluation protocol for unbiased performance assessment. We introduce a grouped leave-one-parent-sequence-out evaluation scheme designed to reduce data leakage caused by overlapping subsequences and to provide a more conservative estimate of generalization. Together, these contributions provide an exploratory framework for task-aware preprocessing under reduced-leakage evaluation.

The remainder of this paper is organized as follows. Section 2 describes the materials and methods, including the preprocessing strategies, feature extraction, reconstruction pipeline, and evaluation protocol. Section 3 presents the experimental results and analysis. Section 4 discusses the findings, limitations, and implications for underwater vision applications. Finally, Section 5 concludes the paper and outlines directions for future work.

2. Materials and Methods

2.1. Overall Framework

Underwater sparse 3D reconstruction is strongly affected by wavelength-dependent attenuation, scattering, reduced contrast, blur, and color distortion in water [4]. Recent studies have shown that image enhancement may influence downstream geometric tasks, but the best preprocessing strategy is often scene-dependent rather than globally fixed [10,11,24]. Motivated by this observation, the present study formulates preprocessing selection as a lightweight machine-learning problem instead of designing a new enhancement network. (Figure 1).

Given an underwater image subsequence s = {I₁, I₂,…, I_N}, a candidate preprocessing pool is defined as

$$M=\{\mathrm{Raw}, \mathrm{CLAHE}, \mathrm{Grayworld}, \mathrm{WB}+\mathrm{CLAHE}\}$$

(1)

Each candidate-transformed subsequence is reconstructed by a sparse Structure-from-Motion pipeline. Let $P(s,m)$ denote the set of sparse 3D points in the final selected COLMAP model reconstructed from subsequence s using preprocessing method m. The primary comparison variable is then defined as:

$$R\left(s,m\right)=\left|P\left(s,m\right)\right|, m\in M$$

(2)

where $\mid \cdot \mid$ denotes set cardinality.

The complete workflow consists of long-sequence splitting, candidate preprocessing generation, sequence-level descriptor extraction, sparse reconstruction, and machine-learning-based preprocessing recommendation. The artificial intelligence component of the framework lies in the learned decision module that predicts which candidate preprocessing strategy is more likely to improve downstream reconstruction.

2.2. Data Sources and Sequence Organization

The experiments were conducted on five underwater parent sequences derived from the Underwater Image Quality Set (UIQS) subset of the Real-world Underwater Image Enhancement (RUIE) benchmark [13]. Each parent sequence contained 726 temporally ordered frames, resulting in a total of 3630 raw underwater images. These sequences were used as the reconstruction benchmark because they preserved the temporal continuity and viewpoint progression required by sparse Structure-from-Motion.

Each parent sequence was initially divided into three overlapping subsequences, yielding 15 subsequences in total. Since four preprocessing candidates were evaluated for each subsequence, the benchmark accounting reported in this paper corresponds to 60 original subsequence–method sparse reconstruction runs. Some weak subsequences were additionally inspected through shorter overlapping windows during reconstruction troubleshooting. However, these refined windows were not treated as new statistical units in the final reporting. Unless otherwise stated, all quantitative counts in this paper therefore continue to refer to the original 15 subsequences.

2.3. Lightweight Preprocessing Candidates

Four lightweight preprocessing candidates were considered in this study: raw input, contrast-limited adaptive histogram equalization (CLAHE), gray-world white balancing, and white-balance-plus-CLAHE. The raw branch leaves the image unchanged. CLAHE was applied to the luminance channel in LAB color space with a clip limit of 2.0 and a tile-grid size of 8 × 8 [25]. Gray-world balancing was used to compensate channel imbalance through global channel scaling. The white-balance-plus-CLAHE variant combined gray-world correction and CLAHE in sequence.

These four methods were selected because they are computationally inexpensive, easy to reproduce, and representative of practical low-cost preprocessing options for underwater reconstruction [11,25].

For clarity, the main symbols used in Equations (1)–(8) are summarized in

Table A1. Main notation used in the proposed framework.

Symbol	Meaning
$s$	Underwater image subsequence
$p\left(s\right)$	Parent sequence of subsequence $s$
$m$	Preprocessing method
$M$	Candidate preprocessing pool
$R(s,m)$	Number of reconstructed sparse 3D points
$J(s,m)$	Number of registered images
$g(s,m)$	Relative gain over raw input
${\tau }_{reg}$	Registration threshold
$I_i$	The $i$-th frame in subsequence $s$
$N$	Number of frames in subsequence $s$
$P(s,m)$	Set of sparse 3D points reconstructed from subsequence $s$ using preprocessing method $m$

.

2.4. Subsequence Generation and Refinement

Each original sequence was first divided into three overlapping subsequences to reduce long-range appearance drift and unstable overlap. The overlap in the initial split was fixed to 40 frames.

A subsequence was regarded as weak when none of the candidate preprocessing methods produced a sufficient number of registered images:

$$\underset{m\in M}{\max } J\left(s,m\right)<{\tau }_{reg}$$

(3)

For difficult clips, shorter overlapping windows were used only as an internal troubleshooting step to inspect whether local overlap could be improved. In this refinement stage, the window length was set to 150 frames, the overlap was set to 30 frames, and the minimum allowed window length was 80 frames. These refined windows were not counted as additional benchmark subsequences in the final reporting.

2.5. Sequence-Level No-Reference Descriptors

For each subsequence–method pair, the first 20 frames were sampled in temporal order. From these sampled frames, a compact set of sequence-level no-reference descriptors was extracted. The final feature vector is expressed as

$$x(s,m)=\left[\overline{UCIQE},\hspace{0.33em}\overline{H},\hspace{0.33em}{\overline{V}}_{lap},\hspace{0.33em}{\overline{C}}_{rms},\hspace{0.33em}{\overline{D}}_{edge},\hspace{0.33em}\overline{Colorfulness},\hspace{0.33em}{\mu }_I,\hspace{0.33em}{\sigma }_I\right]$$

(4)

These descriptors summarize underwater color distortion, entropy, blur, contrast, edge structure, color richness, and intensity variation [13,14,20,24]. Since the goal of the model is not to generate images but to recommend a preprocessing policy, compact statistical descriptors were preferred over high-dimensional image embeddings.

2.6. Sparse 3D Reconstruction

For each subsequence–method pair, an independent sparse reconstruction was performed using COLMAP [12]. The reconstruction process included feature extraction, feature matching, geometric verification, and incremental mapping. In addition to the number of reconstructed sparse 3D points R(s, m), the number of registered images, the mean track length, and the mean reprojection error were also recorded.

If multiple sparse models were produced for the same subsequence–method pair, the final model was selected according to a lexicographic rule that prioritized more reconstructed points, then more registered images, and finally smaller reprojection error.

2.7. Artificial-Intelligence-Based Preprocessing Selection

This section is the core artificial intelligence component of the proposed framework. Instead of learning image appearance, the model learns preprocessing utility for downstream reconstruction.

For each subsequence s, let the raw reconstruction score be R(s, raw). The learning target is defined as the relative gain of candidate method m over raw input:

$$g(s,m)={\displaystyle \frac{R(s,m)-R(s,\mathrm{raw})}{\max (R(s,\mathrm{raw}),1)}}$$

(5)

This formulation reduces the dominance of easy subsequences with unusually large point counts and directly reflects the practical decision question of whether a candidate preprocessing method is likely to improve reconstruction relative to the raw input.

Two lightweight regressors were considered, namely ridge regression and random forest regression. Within each outer training fold, an inner grouped cross-validation was used to select the better regressor according to the mean absolute error. For a test subsequence s, the trained model predicts the gain of each candidate method, and the recommended preprocessing method is defined as

$$\widehat{m}(s)=\underset{m\in \mathcal{M}}{\arg \max }\widehat{g}(s,m)$$

(6)

In this way, machine learning is integrated with underwater sparse reconstruction at the preprocessing-policy level: sequence-level descriptors are used as input, actual reconstruction outcomes are used as supervision, and the final output is a preprocessing decision for the target subsequence.

Ridge regression and random forest regression were selected as two complementary lightweight regressors under small-sample grouped evaluation. Ridge provides a stable linear estimator when handcrafted sequence-level descriptors are correlated, whereas random forest can capture nonlinear interactions among descriptors without heavy hyperparameter tuning. These models were prioritized because the present study focused on practical preprocessing selection under limited grouped data rather than on model-size scaling. Ridge regression also served as a regularized linear reference under correlated handcrafted descriptors. Additional simpler baselines were not separately emphasized, and more complex models were intentionally not pursued because the grouped dataset was small and the goal of the study was lightweight, deployable preprocessing selection rather than model-size scaling.

2.8. Grouped Leave-One-Parent-Sequence-Out Evaluation

Let p(s) denote the parent sequence of subsequence s. To avoid train–test leakage caused by overlapping clips extracted from the same original sequence, evaluation was performed using a grouped leave-one-parent-sequence-out protocol:

$$S_{\mathrm{test}}^{(f)}=\{s \mid p(s)=P_f\},\hspace{1em}{S}_{\mathrm{train}}^{(f)}=\{s \mid p(s)\ne P_f\}$$

(7)

All clips originating from the same parent sequence were therefore held out together. This grouped protocol is intended to reduce train–test leakage caused by overlapping subsequences and to provide a more conservative estimate of generalization.

2.9. Fair Fixed Baseline and Oracle

For each outer fold, a fair fixed baseline was defined by selecting one single preprocessing method using only the training fold. Let the average training gain of method m be

$${\overline{g}}_{\mathrm{train}}(m)={\displaystyle \frac{1}{|S_{\mathrm{train}}^{(f)}|}}{\displaystyle \sum _{s\in S_{\mathrm{train}}^{(f)}}g}(s,m)$$

(8)

The fixed baseline used the method with the highest average training gain. In addition, an oracle upper bound was defined by selecting the true best-performing preprocessing method for each test subsequence. The final grouped comparison therefore included four outputs: raw input, the fair fixed baseline, the machine-learning-based selected method, and the oracle upper bound.

2.10. Statistical Analysis

The primary comparison variable was the number of reconstructed sparse 3D points. Because reconstruction outcomes were highly heterogeneous across subsequences, both mean and median values were reported in the grouped comparison. In addition, win/tie/loss counts were computed to compare the machine-learning selector with the fixed baseline on a per-subsequence basis. In addition, reconstruction pass rate under the predefined registration threshold was reported as a complementary viability metric across the full benchmark.

For model selection within the training folds, mean absolute error was used as the optimization criterion. For the final grouped evaluation, descriptive statistics were reported for the raw input, the fair fixed baseline, the selected method, and the oracle upper bound. No significance test was used because the final grouped comparison set was small and the aim of the study was exploratory task-oriented evaluation rather than formal population inference.

2.11. Implementation Details

All preprocessing, feature extraction, sequence splitting, grouped evaluation, and result aggregation were performed in Python 3.10. Sparse reconstruction was carried out using the command-line interface of COLMAP [12]. Sequential matching was adopted as the default matching strategy because the input data consisted of temporally ordered underwater subsequences. The camera model was set to a single shared simple radial model across images within the same reconstruction task.

The experimental dataset comprised five parent sequences derived from the UIQS subset of RUIE [13], with 726 frames in each sequence and 3630 frames in total. The preprocessing parameters were fixed for all experiments: CLAHE clip limit = 2.0, CLAHE tile-grid size = 8 × 8, initial split overlap = 40 frames, refined split window length = 150 frames, refined split overlap = 30 frames, sampled frames per subsequence–method pair = 20, and registered-image threshold = 10.

For the artificial intelligence module, ridge regression and random forest regression were used as lightweight candidate models. The ridge model used a regularization coefficient of 1.0, while the random forest used 300 trees. These choices were intentionally kept simple in order to emphasize preprocessing selection under low implementation complexity rather than model over-optimization.

The refined split parameters were used only for troubleshooting weak clips and did not alter the benchmark accounting based on the original 15 subsequences and 60 subsequence–method runs.

Runtime statistics were measured on the same hardware environment for candidate preprocessing generation, descriptor extraction, selector inference, and sparse COLMAP reconstruction.

3. Results

3.1. Main Quantitative Set and Failure Cases

After threshold-based filtering, only six of the original 15 subsequences remained suitable for the main grouped comparison, while the remaining nine were categorized as failure cases. This result itself is important because it shows that underwater sparse reconstruction is constrained not only by the choice of preprocessing, but also by the inherent viability of the sequence. The benchmark accounting reported in this paper continues to refer to the original five parent sequences, 3630 frames, 15 subsequences, and 60 subsequence–method reconstruction runs.

The subset assignment by parent sequence is summarized in

Table 1. Number of subsequences assigned to the main quantitative set and the failure-case set for each parent sequence.

Parent Sequence	Main Quantitative	Failure Case	Total
seq_01	3	0	3
seq_02	2	1	3
seq_03	1	2	3
seq_04	0	3	3
seq_05	0	3	3

and visualized in Figure 2. Among the five parent sequences, seq_01 contributed three subsequences to the main quantitative set, seq_02 contributed two, and seq_03 contributed one, whereas all subsequences from seq_04 and seq_05 remained in the failure-case set under the predefined registration threshold. Therefore, the final grouped comparison reported below was effectively supported by three parent sequences.

Table 2. Full-benchmark subsequence viability and failure breakdown across the original 15 subsequences.

Parent Sequence	Original Subsequence	Best Viability Method	Best Registered Images	Best Point Method	Best Sparse Points	Pass/Fail	Best Troubleshooting Window
seq_01	seq_01_part01	CLAHE	136	CLAHE	19,398	Pass	—
seq_01	seq_01_part02	CLAHE	65	CLAHE	10,338	Pass	—
seq_01	seq_01_part03	WB + CLAHE	58	WB + CLAHE	8142	Pass	—
seq_02	seq_02_part01	WB + CLAHE	6	WB + CLAHE	1294	Fail	seq_02_part01_r01
seq_02	seq_02_part02	CLAHE	10	CLAHE	805	Pass	—
seq_02	seq_02_part03	CLAHE	27	CLAHE	3113	Pass	—
seq_03	seq_03_part01	WB + CLAHE	6	WB + CLAHE	1542	Fail	seq_03_part01_r01
seq_03	seq_03_part02	WB + CLAHE	11	Gray-world	875	Pass	—
seq_03	seq_03_part03	Gray-world	3	CLAHE	603	Fail	—
seq_04	seq_04_part01	CLAHE	4	CLAHE	359	Fail	seq_04_part01_r02
seq_04	seq_04_part02	CLAHE	5	CLAHE	501	Fail	seq_04_part02_r02
seq_04	seq_04_part03	WB + CLAHE	5	WB + CLAHE	298	Fail	seq_04_part03_r01
seq_05	seq_05_part01	WB + CLAHE	6	WB + CLAHE	1350	Fail	seq_05_part01_r03
seq_05	seq_05_part02	WB + CLAHE	6	WB + CLAHE	1348	Fail	seq_05_part02_r01
seq_05	seq_05_part03	CLAHE	3	CLAHE	276	Fail	seq_05_part03_r01

Note: A subsequence was labeled as Pass if at least one preprocessing candidate achieved 10 or more registered images; otherwise, it was labeled as Fail. “Best viability method” denotes the candidate with the highest registered-image count, with ties broken by higher sparse-point count and then by lower reprojection error. For weak subsequences that were additionally inspected through shorter troubleshooting windows, the reported troubleshooting-window identifiers are descriptive only and do not alter the benchmark accounting based on the original 15 subsequences.

reports full-benchmark viability across the original 15 subsequences. Overall, 6 subsequences passed the predefined registration criterion, whereas 9 failed under all candidate preprocessing strategies. The best registered-image counts ranged from 10 to 136 in the passing subset and from 3 to 6 in the failure subset. By contrast, best sparse-point counts remained nonzero even in failed cases, ranging from 276 to 1542. The failed subsequences were not excluded because they contained zero sparse points, but because none of the candidate preprocessing strategies reached the minimum registration criterion. In practice, these failures were characterized by persistently insufficient image registration (best registered-image counts of only 3–6) and weak, fragmented reconstructions even after local troubleshooting. This confirms that sequence-level reconstruction viability cannot be judged by sparse-point count alone and supports reporting pass/fail statistics alongside the main geometric comparison.

3.2. Overall Comparison of Raw, Fixed Baseline, Selected Method, and Oracle

In the grouped evaluation set, the raw input yielded a mean of 5420.3 reconstructed sparse points. The machine-learning-based selector increased this value to 6518.2, corresponding to a relative improvement of 20.3% over raw input. The fair fixed baseline achieved 6582.0 reconstructed sparse points on average, corresponding to a 21.4% improvement over raw input. The oracle upper bound reached 7111.8 reconstructed sparse points on average, which was 31.2% higher than raw input.

These results indicate that preprocessing selection can substantially influence sparse reconstruction quality. At the same time, they also show that a strong fixed baseline remained highly competitive under the grouped evaluation protocol. The per-subsequence comparison is listed in

Table 3. Per-subsequence grouped comparison of raw input, the fair fixed baseline, the machine-learning-based selected method, and the oracle upper bound.

Subsequence	Raw	Fixed Baseline	Selected	Oracle
seq_01_part01	15,303	CLAHE 19,398	WB + CLAHE 16,565	CLAHE 19,398
seq_01_part02	8933	CLAHE 10,338	WB + CLAHE 10,022	CLAHE 10,338
seq_01_part03	5560	CLAHE 5428	WB + CLAHE 8142	WB + CLAHE 8142
seq_02_part02	379	WB + CLAHE 694	Gray-world 739	CLAHE 805
seq_02_part03	1675	WB + CLAHE 2948	WB + CLAHE 2948	CLAHE 3113
seq_03_part02	672	CLAHE 686	WB + CLAHE 693	Gray-world 875

, the summary statistics are reported in

Table 4. Summary statistics of reconstructed sparse points in grouped evaluation.

Strategy	Mean	Median	Improvement Over Raw (%)
Raw	5420.3	3617.5	—
Fixed baseline	6582.0	4188.0	21.4
Selected	6518.2	5545.0	20.3
Oracle	7111.8	5627.5	31.2

, and the grouped comparison is visualized in Figure 3.

3.3. Secondary Geometric Indicators and Accuracy Trade-Offs

Table 5. Secondary geometric indicators in grouped evaluation.

Strategy	Mean Sparse Points	Mean Registered Images	Mean Track Length	Mean Reprojection Error	Pass Rate (%)
Raw	5420.3	45.5	4.289	0.748	66.7
Fixed baseline	6582.0	48.5	4.758	0.917	100.0
Selected	6518.2	47.5	4.620	0.838	83.3
Oracle	7111.8	49.7	4.580	0.875	83.3

Note: Mean sparse points remained the primary comparison variable. The auxiliary indicators reported here were used to examine whether increases in reconstructed sparse points were accompanied by changes in geometric consistency. Pass rate denotes the percentage of subsequences whose selected strategy achieved 10 or more registered images. Because the oracle was defined with respect to sparse-point count rather than registration viability, its pass rate is not necessarily 100%.

reports secondary geometric indicators for the grouped evaluation set. Although sparse-point count remained the primary comparison variable, the auxiliary results show that preprocessing also changed registration behavior and geometric consistency. Compared with raw input, the selector increased sparse-point count, mean registered images, and mean track length, but with a modest increase in mean reprojection error (0.838 vs. 0.748). Compared with the fixed baseline, however, the selector achieved a lower mean reprojection error (0.838 vs. 0.917) while remaining close to it in sparse-point count. These results indicate a mixed geometry trade-off rather than a systematic degradation of geometric accuracy. At the same time, pass rates differed across strategies, confirming that reconstruction viability cannot be assessed from sparse-point count alone.

3.4. Per-Subsequence Behavior

A more detailed analysis reveals that the selector improved over the raw input on all six evaluated subsequences. In comparison with the fair fixed baseline, the selector outperformed the fixed strategy on three subsequences, tied it on one subsequence, and underperformed it on two subsequences. This indicates that the selector improved over raw input, but was not yet consistently superior to the strongest fixed preprocessing choice.

The median number of reconstructed points further supports this observation. The selector achieved a median of 5545.0 reconstructed points, whereas the fixed baseline achieved 4188.0. Hence, although the fixed baseline slightly exceeded the selector in mean performance, the selector remained competitive on typical clips.

The oracle method also varied across subsequences. In the evaluated set, the oracle was CLAHE on four subsequences, white-balance-plus-CLAHE on one subsequence, and gray-world white balancing on one subsequence. This variability directly supports the main hypothesis of the study: the best preprocessing strategy is not universally fixed, but depends on the visual and geometric characteristics of the subsequence.

3.5. Method Frequency and Selection Behavior

The method frequencies in grouped evaluation are shown in Figure 4. The fair fixed baseline selected CLAHE in four cases and white-balance-plus-CLAHE in two cases. In contrast, the machine-learning selector preferred white-balance-plus-CLAHE in five cases and gray-world white balancing in one case. The oracle upper bound was still dominated by CLAHE, but not exclusively so.

These counts indicate that CLAHE acted as a strong robust baseline, whereas the selector behaved more aggressively and favored white-balance-plus-CLAHE more often. This difference helps explain why the selector exceeded the fixed baseline in some clips while remaining slightly lower on mean performance.

3.6. Runtime Overhead of the Selection Module

Table 6. Runtime breakdown of the selection framework.

Stage	Mean Time per Subsequence (s)	Share of One Selected Sparse Reconstruction (%)
Candidate preprocessing generation	2.4	8.0
Descriptor extraction	0.8	2.7
Selector inference	0.001	0.0
Total selector-related overhead	3.201	10.7
One sparse COLMAP reconstruction	30.0	100.0

reports the runtime overhead of the proposed selection framework. Candidate preprocessing generation, descriptor extraction, and selector inference together required 3.201 s per subsequence on average, corresponding to 10.7% of the average sparse COLMAP reconstruction time (30.0 s). Among these components, candidate preprocessing generation accounted for the largest share of the selector-related cost, whereas model inference itself was negligible. By comparison, exhaustively reconstructing all four preprocessing candidates was estimated to require approximately 120.0 s per subsequence, corresponding to about 400.0% of one selected sparse reconstruction. These results support the characterization of the proposed selector as lightweight relative to the overall sparse reconstruction pipeline.

4. Discussion

The results support three main observations. First, underwater sparse 3D reconstruction is strongly influenced by lightweight preprocessing, confirming that radiometric changes can materially affect downstream geometry [10,11]. This finding reinforces the need to evaluate enhancement methods in a task-aware manner rather than relying solely on perceptual quality improvements.

Second, the optimal preprocessing strategy is inherently scene-dependent. Although CLAHE emerged as a strong baseline, the oracle method still varied across subsequences, indicating that no single enhancement strategy is universally optimal. This observation is consistent with recent task-oriented studies showing that visual enhancement and geometric utility may not coincide [11,24]. From a practical perspective, this variability motivates adaptive selection mechanisms rather than fixed preprocessing pipelines.

Third, the evaluation protocol plays a critical role in estimating generalization performance. By holding out all subsequences from the same parent sequence together, the grouped setting is intended to reduce overlap-induced leakage and provide a more conservative estimate of generalization. Under this stricter protocol, the proposed selector consistently improved over raw input across all evaluated subsequences. At the same time, it did not outperform the strongest fixed baseline in mean performance. This is an important limitation of the current study. The present results therefore provide preliminary evidence that adaptive preprocessing can improve robustness relative to raw input while remaining competitive with a strong fixed baseline.

The study also has several limitations. The grouped comparison set was relatively small, and only a limited number of parent sequences remained quantitatively valid after thresholding, which may restrict the statistical strength of the conclusions. In particular, because the final grouped comparison was effectively supported by only three parent sequences, the current evidence on generalization should be interpreted as preliminary. In addition, the feature representation was intentionally lightweight and relied on sequence-level no-reference descriptors rather than learned embeddings or temporal representations. While this design supports practical deployment, it may limit predictive accuracy in more complex scenarios.

Future work will expand the diversity and scale of parent sequences, investigate richer task-aware descriptors and ranking-based learning objectives, and extend the same decision framework to dense reconstruction, SLAM stability, and broader underwater mapping tasks. These directions will further clarify the role of adaptive preprocessing in improving the reliability of underwater vision systems.

5. Conclusions

This paper presented a machine-learning-based preprocessing selection framework for underwater sparse 3D reconstruction. Instead of proposing a new enhancement network, the study formulated preprocessing selection as a lightweight task-aware decision problem. Four candidate preprocessing strategies were evaluated, including raw input, CLAHE, gray-world white balancing, and white-balance-plus-CLAHE. Sequence-level no-reference descriptors were used to predict reconstruction gains, and a grouped leave-one-parent-sequence-out protocol was adopted to reduce train–test leakage and provide a more conservative estimate of generalization.

The experimental results indicate that underwater sparse reconstruction is highly sensitive to preprocessing choice. Relative to raw input, the proposed selector improved the mean number of reconstructed sparse points from 5420.3 to 6518.2, remaining competitive with a strong fixed baseline (6582.0) and below the oracle upper bound (7111.8). The selector consistently improved over raw input across all evaluated subsequences, but it did not surpass the strongest fixed baseline in mean performance.

Overall, this study suggests three main insights: (1) preprocessing should be treated as a task-aware decision problem rather than a fixed enhancement step; (2) grouped evaluation is useful for reducing overlap-induced leakage and providing a more conservative estimate of generalization; and (3) lightweight machine-learning-based selection is a promising exploratory approach for improving reconstruction robustness relative to raw input under heterogeneous underwater conditions.

These findings provide preliminary evidence on the role of preprocessing in underwater vision systems and offer task-oriented guidance for studying enhancement strategies in real-world 3D reconstruction pipelines.