Beyond Aesthetics: Quantifying Information Loss in Turbid Scenes

Conference: ECCV 2026
Paper: Official paper page · Paper PDF
Project: https://vap.aau.dk/pcd/
Area: Segmentation (instance segmentation benchmarks and underwater information assessment)
Keywords: Underwater turbidity, phase congruency, structural information, no-reference metrics, instance segmentation

TL;DR

The paper introduces the controlled TUB dataset and Phase-Congruency Delentropy (PCD) to distinguish unattractive underwater images from images that have lost task-relevant structure; PCD correlates most strongly with segmentation performance in the reported real and synthetic experiments, but is not a universal measure of recognizability across arbitrary scenes.

Background & Motivation

Underwater degradation changes more than appearance. Absorption alters brightness and color, backscatter reduces contrast, and stronger scattering can spread boundaries until neighboring regions become indistinguishable. Image-enhancement evaluations often collapse these effects into a single notion of poor quality, without asking which information an instance segmentation model actually needs. A low-contrast image may still preserve the locations of object boundaries, whereas color correction cannot recover a boundary that has genuinely disappeared.

Testing this distinction requires reliable labels under conditions where humans cannot confidently draw masks. Annotating only visible objects can quietly remove the hardest instances from evaluation, while relying on synthetic images risks equating color shifts and contrast reduction with real scattering-induced information loss. The paper therefore needs both a controlled dataset with consistent ground truth across degradation levels and a structural measure that is not dominated by contrast changes. Its contribution is an evaluation foundation, not a new segmentation architecture.

The authors combine phase congruency with delentropy and compare the resulting measure with downstream performance under controlled scene structure. Core idea: preserve trustworthy labels through static real-turbidity sequences, then measure multiscale phase structure rather than visual attractiveness to determine which degradations actually matter for instance segmentation.

Method

Overall Architecture

The study comprises controlled TUB acquisition, PCD structural measurement, and real–synthetic task validation. TUB supplies images and instance masks for fixed scenes at different turbidity levels; PCD takes one image and returns a scalar without requiring a clear reference image. Validation relates that scalar to segmentation AP50 on the corresponding images. PCD neither enhances the image nor generates the segmentation masks.

Two experimental tracks must be kept separate. The real-only baseline changes the turbidity subsets used for training and evaluation to assess adaptation and remaining performance loss. The metric evaluation instead trains on a mixture of real images and two synthetic degradation families, then compares per-image AP50 with image-quality and complexity measures. The first track establishes task difficulty; the second asks which metric follows that difficulty. Because this is a dataset, metric, and evaluation study rather than a multicomponent prediction network, a model-inference flowchart would be misleading.

Key Designs

1. Controlled TUB acquisition: retaining ground truth when targets become invisible

Four GoPro11 cameras surround a cylindrical tank, with objects attached to a LEGO baseplate. Natural and manufactured items, including rocks and trash, are rearranged between scenes, but both cameras and objects remain fixed throughout each turbidity sequence. Consequently, changes within a viewpoint primarily reflect the medium rather than a moving target or changing composition. Each scene starts with clear water and proceeds through ten stages using measured additions of oat milk. A nephelometer measures three water samples at every stage. Since identical nominal additions do not ensure identical turbidity across scenes, images are grouped by measured NTU rather than addition index. The paper defines low turbidity as 0–10 NTU, medium as 10–20 NTU, and high as 20+ NTU. There are 33 scenes with 40 images each, yielding 1,320 images.

Masks are drawn on clear images and propagated within each viewpoint to all corresponding turbid images. The static geometry makes reference masks available even when a human could no longer delineate the boundaries reliably, producing over 16,000 instance masks. All objects share one class, so the benchmark primarily measures instance localization and delineation rather than semantic recognition across object categories. This choice also matters for interpreting scores: a model is penalized for missing a fully obscured object, not merely evaluated on whichever objects remain visibly annotatable. Multiple cameras introduce viewpoint and occlusion variation, but their images are not independent scenes. The evaluation therefore splits at scene level rather than randomly distributing individual images across training and validation.

2. PCD structural measurement: reducing contrast dependence before measuring complexity

Conventional delentropy measures the joint entropy of horizontal and vertical image-gradient components rather than the entropy of raw pixel intensities. This captures spatial variation, but gradients computed with fixed kernels still shrink when contrast decreases, even if boundary locations remain intact. A fixed single scale also struggles to describe both fine edges and broad, blurred contours. PCD retains the joint-statistics idea while replacing the representation being measured; it does not simply reweight existing quality scores.

Phase congruency identifies positions where local frequency components align in phase, which tend to correspond to salient structure. The paper uses Kovesi's implementation, with wavelets derived from Gabor functions across multiple scales and noise suppression during estimation. Responses range from 0 to 1 and emphasize structural alignment rather than directly treating raw intensity differences as edge strength. Following the two-dimensional delentropy construction, the authors compute two orthogonal orientations, producing two phase-congruency response maps in place of conventional gradient components. A joint distribution of the paired responses is then used to compute joint entropy, giving Phase-Congruency Delentropy, or PCD. A higher PCD indicates richer structural responses in this representation, not necessarily more objects or a guaranteed higher segmentation score.

This substitution helps avoid interpreting contrast reduction or discoloration alone as severe structural loss when boundaries remain available. When scattering spreads or removes boundaries, the multiscale responses change in a way that can better reflect the segmentation challenge. However, phase representations do not recreate information from nothing: once neighboring regions collapse to identical pixel values, their distinction may genuinely be lost. An image can also retain sharp nearby edges while distant objects disappear completely, producing an overly optimistic structural score. The cached rendering of Section 4, Equation (1), has damaged symbols and summation indices, so this note uses the prose-supported joint-entropy definition rather than reconstructing an uncertain equation. The main text also does not fully specify discretization of the joint distribution; exact reproduction requires the project implementation rather than an invented binning rule.

3. Real–synthetic task validation: testing whether structure follows task utility

The authors generate two synthetic families from the lowest-turbidity images, producing ten variants per input for each method. Synth1 uses an underwater image-formation model that primarily introduces distance-dependent contrast loss. Synth2 additionally includes blur and spatially varying inhomogeneity noise, testing degradations that more directly damage boundaries. ZoeDepth estimates depth, its outputs are randomly rescaled within 0.2–4 meters, and attenuation coefficients are sampled from a distribution derived from Jerlov water types. This represents synthesis with unknown geometry and medium parameters, not a physical inversion calibrated to each tank scene. Synthetic images retain the original masks and are combined with real images, tripling the dataset volume.

Three segmentation models are trained on this combined dataset and evaluated using per-image AP50. Their results are compared with Entropy, Delentropy, NIQE, UIQM, UCIQE, and the reference-based SSIM and PSNR, which use the lowest-turbidity image as reference. None of these metrics has its statistics or coefficients fine-tuned on TUB, and PCD is not a learned regressor trained to predict model AP50. Thus, the relationship being tested is between an independently computed measure and task performance, rather than an explicitly supervised score predictor. The segmentation models still require masks, and validating correlation requires task outcomes; “no-reference” only describes the inputs needed to calculate PCD itself. Synth1 frequently depresses conventional quality scores while preserving accurate predictions, exposing the appearance–utility mismatch that motivates the study.

Loss & Training

PCD requires no training, and the paper introduces neither a new segmentation loss nor a modified prediction architecture. Real TUB experiments use 25 training scenes and 8 validation scenes, with low-, medium-, high-, and all-turbidity subsets. The same object may occur in both splits, but not in the same spatial arrangement or position; this is a scene split, not an object-identity-disjoint split. Table 1 uses ResNet101-FPN for MaskRCNN, YOLO11m-seg for YOLOv11, and SwinB for Mask2Former. Learning rates are respectively \(10^{-5}\), \(10^{-3}\), and \(10^{-4}\), with batch sizes of 8, 32, and 8. Although the table lists 48 epochs for MaskRCNN and Mask2Former, its footnote specifies 6,000 detectron2 steps; YOLOv11 uses 50 epochs. Absolute differences therefore include backbone and optimization choices rather than establishing an equal-budget architectural advantage.

For the main PCD experiments, images are downsampled by a factor of 0.5, with seven wavelet scales, minimum wavelength 3, and multiplication factor 2.1. The sensitivity analysis separately identifies six scales and minimum wavelength 2 as more stable overall; that finding is not the configuration used for the main results. Increasing scale count incorporates lower-frequency structure and reduces the relative contribution of fine-scale content, so more scales are not automatically better.

Key Experimental Results

Main Results

The selected results below come from Figure 4, using only real TUB images and training each model on all turbidity levels. Evaluation uses the 8 held-out scenes; AP50 is instance segmentation average precision at IoU threshold 0.5, reported on a percentage scale, with higher values better. These are benchmark baselines, not improvements obtained by adding PCD to a model: PCD does not participate in training or prediction.

Model	All validation images AP50 ↑	Low turbidity AP50 ↑	Medium turbidity AP50 ↑	High turbidity AP50 ↑
MaskRCNN	41.8	64.6	46.6	19.8
YOLOv11	50.8	85.9	53.0	17.9
Mask2Former	51.8	87.6	55.5	17.1

Mask2Former leads overall but not on the high-turbidity subset, so its aggregate advantage should not be generalized to every operating condition. The “All” score is evaluated over all validation images, not obtained by taking the arithmetic mean of the three turbidity-specific columns.

Ablation Study

There is no removal-based ablation of newly proposed network components. The following genuine training-distribution analysis, also from Figure 4, fixes Mask2Former and changes its real-data training subset. Evaluation retains the same 8 held-out scenes; higher AP50 is better throughout.

Training data	All validation images AP50 ↑	Low turbidity AP50 ↑	Medium turbidity AP50 ↑	High turbidity AP50 ↑
Low only	46.8	89.4	51.2	8.5
Medium only	48.4	84.6	52.9	14.5
High only	42.4	68.9	48.9	17.1
All turbidity levels	51.8	87.6	55.5	17.1

Moving from low-only to all-turbidity training raises high-turbidity AP50 from 8.5 to 17.1, an increase of 8.6 percentage points, while reducing low-turbidity AP50 from 89.4 to 87.6, a decrease of 1.8 points. Changing the training subset also changes available image count, so this is not a sample-count-controlled single-variable ablation. PCD scale and resolution sensitivity is analyzed separately in Figure 9; values that cannot be read precisely from those curves are not converted into a fabricated numerical table here.

Key Findings

• Exposure to degradation helps but does not remove information loss. Mixed-turbidity training improves difficult conditions, yet the highest-turbidity subset remains challenging.
• Appearance is not the only relevant criterion. Figures 6–8 show that Synth1 can preserve useful boundaries despite low contrast, separating conventional quality scores from task performance.
• PCD is a structural proxy, not a perfect predictor. The authors report the strongest correlations on real and combined data, while also identifying outliers caused by fully obscured objects.
• Evidence boundary: the local text does not preserve exact correlation coefficients from Figures 6–7, and the prose does not explicitly identify the correlation statistic. This note therefore reports the supported relative finding without inventing values or labeling them Pearson correlations.

Highlights & Insights

• The annotation protocol addresses an extreme-condition evaluation blind spot. Propagating clear-image masks prevents invisible objects from silently disappearing from ground truth, although the guarantee depends on static geometry.
• Changing the representation targets the actual confound. PCD retains joint entropy but replaces contrast-sensitive fixed gradients with phase structure, separating edge presence from raw brightness differences more directly.
• Synthetic degradation needs a task-based reference. Poor appearance alongside high AP50 can reveal a failure to reproduce structural damage; PCD can help diagnose that mismatch rather than certify physical realism by itself.

Limitations & Future Work

• Author-stated scene dependence: no-reference information measures respond to inherent scene detail. A featureless clear-water region and a richly textured turbid scene cannot reliably be ranked by PCD alone; comparable structural complexity is an important condition of the TUB experiments.
• Author-discussed occlusion failure: sharp surviving edges can sustain a high score while segmentation is penalized for completely invisible objects. Global structural richness is not the same as recoverability of every instance.
• Reader assessment of scope: oat-milk tank scenes, stationary objects, and single-class masks do not establish generalization to natural particles, moving animals, complex illumination, or multiclass recognition in open water.
• Reader assessment of statistics: the scene count is small, and the main text does not report repeated-run uncertainty or significance for the selected results. Correlation is not a causal guarantee connecting a structural score to performance.
• Next validation priority: broader real-world acquisition and repeated experiments across different inherent scene complexities should precede stronger claims about using PCD for data selection or synthetic-pipeline comparison.

Related Work & Insights

• vs conventional Delentropy: both measure joint entropy of two-dimensional structural responses. PCD uses multiscale phase congruency instead of fixed-kernel gradients that are directly affected by contrast.
• vs UIQM, UCIQE, and NIQE: these emphasize perceptual quality or image statistics, whereas the present study targets structure associated with instance segmentation utility. Lower task correlation does not invalidate them for their original image-quality purposes.
• vs underwater image enhancement: the paper neither generates clearer images nor establishes the effectiveness of an enhancement network. It supplies data and a measure for checking whether enhancement or synthesis preserves task information.
• vs MaskRCNN, YOLOv11, and Mask2Former: these are probe models used to measure task utility, not competing methods that PCD replaces; the contributions operate at different levels.

Rating

• Novelty: 4/5 — Phase congruency and joint entropy are established, but their combination with controlled extreme-turbidity data addresses a well-defined evaluation gap.
• Experimental Thoroughness: 3/5 — Three models, real and synthetic conditions, and parameter analysis provide useful coverage; scene scale and open-water validation remain limited.
• Writing Quality: 4/5 — The distinction between appearance and structural utility is clear, while exact reproduction still needs implementation and statistical details.
• Value: 4/5 — Useful for underwater perception evaluation and synthetic-degradation diagnosis, without establishing a universal replacement for image-quality measures.