Judging the quality of a 3D image is not the same as judging two flat images. When only one eye's view is blurred or compressed, viewers do not simply average the two views. Their brains suppress one view or let the views compete, a process called binocular rivalry, and the result depends on which eye is cleaner and what kind of damage was done. Stereoscopic quality assessment is the task of predicting human ratings of such left/right image pairs. Its central difficulty is that labelled data are tiny, typically a few hundred pairs per database, while the behavior it must learn is nonlinear. Current deep models handle this in one of two ways. They pretrain on 2D images and fine-tune (PAD-Net), or they train a large pretrained backbone on everything at once (iSQoE). Neither ever shows the model the binocular part of the problem in an organized way.
We propose to pretrain a stereo quality model on unlimited synthetic stereopairs whose damage we control, using an ordering we call a binocular curriculum. The model first learns to rank symmetric damage, where both eyes degrade together and 2D intuition mostly holds. One-eye (asymmetric) damage is then mixed in, and finally geometric damage between the views. Throughout, the earlier stages stay in the training mix so they are not forgotten. The supervision costs nothing: within a pristine source pair, adding damage to either view while holding the other fixed can only make the image worse, so the model learns from these guaranteed orderings. The claim is that the ordering, not just the synthetic data, is what makes the model generalize on asymmetric distortions and across databases.
Core Claim
Curriculum learning means training on examples in order of increasing difficulty. Its gains reported so far come from settings where difficulty has one axis, such as JPEG compression from mild to severe in AR-CNN. Its failures come from orderings that discard earlier tasks: a strictly sequential linear→quadratic→cubic schedule forgets, while a mixed schedule that keeps easy tasks present succeeds (multi-task function-class study). That study is the source of our method choice, since the stereo curriculum will be mixed rather than sequential.
Stereo quality has at least two separable axes, and prior work motivates treating them separately. Zhou and Wang show that depth quality and 2D image quality can vary independently. Fang et al. and PAD-Net show that asymmetric distortion breaks symmetric averaging.
Our hypothesis is specific. A model pretrained with a mixed curriculum along the monocular → interocular-asymmetry → geometric axis should outperform the same model pretrained on the identical synthetic data in shuffled order. The gap should appear on the asymmetric subset of LIVE 3D Phase II and in cross-database transfer, not necessarily on symmetric distortions. If the shuffled model matches it, the claim is false, and the useful contribution reduces to the synthetic ranking data itself.
Research Program
The backbone follows iSQoE: DINOv2 features for each view, joined by cross-attention layers so the views interact before a scalar score is produced. We keep this architecture fixed so that only training order varies.
Synthetic pairs come from pristine stereo images (Holopix50k and Middlebury pairs). We apply the five LIVE distortion families at graded severities: white noise, Gaussian blur, fast fading, JPEG and JPEG2000. We also apply SCOPE-style spatial warps and misalignments between views. Pairwise ranking labels are drawn only from orderings guaranteed by construction. A pair is always worse than the same pair with extra damage on one view and the other view unchanged. Comparisons between, say, symmetric and asymmetric versions are never labelled, because their human ordering is exactly what is unknown.
Training uses a two-alternative ranking loss, the same form iSQoE uses on human preferences. The curriculum has three stages. Stage A uses symmetric-only ranking pairs. Stage B mixes in one-eye monotonic pairs. Stage C adds geometric pairs. The proportion of each newly added type rises linearly across its stage while earlier types keep a floor share, which is the mixed schedule. After pretraining, every variant is fine-tuned on human scores with an identical protocol.
We also run a diagnostic analysis. For asymmetric test pairs, we perturb the cleaner view and the dirtier view separately and measure the change in predicted score. A model that has learned rivalry should be more sensitive to the cleaner view, in line with the dominance weighting PAD-Net builds in by hand. This tests whether the curriculum changes what the model represents, not just its correlation scores.
Baselines
The naive reference is the DINOv2 cross-attention model fine-tuned directly on each human-labelled database with no synthetic pretraining. It shows how much any pretraining helps.
The central ablation is uniform synthetic pretraining. It uses the exact same synthetic pairs, total steps, optimizer and tuning budget as the curriculum, presented shuffled. Beating it is the only result that supports the core claim. We require a consistent improvement across seeds on the LIVE Phase II asymmetric subset and on cross-database transfer; a win on one split alone does not count.
A sequential-curriculum ablation runs the same three stages without the replay floor. It tests whether forgetting, as reported in the function-class study, also appears here.
A generic self-paced baseline applies Self-Paced Learning to the same data. Self-paced learning includes only pairs whose current loss falls below a threshold, and the threshold rises over training. This baseline tests whether the stereo-specific difficulty axis matters or whether loss-based pacing suffices.
An asymmetric-oversampling baseline keeps uniform order but upweights one-eye pairs. It isolates ordering from mere exposure.
Against published systems, the strongest is PAD-Net, which reports SROCC 0.967 on LIVE 3D Phase II, against 0.947 for StereoQA-Net. On LIVE Phase I, Fang et al. report PLCC 0.951. On SOLID, SOIQE reports SROCC 0.924 (SOIQE), and Zhou and Wang report depth-quality SROCC 0.9299. On SCOPE, iSQoE reports 73.1% total accuracy and 84.8% on unanimous pairs. For Waterloo-IVC we found no reported result on Lacuna.
Because Phase II SROCC is already near ceiling, matching 0.967 while improving the asymmetric subset and transfer would count as success. We will retrain PAD-Net under our splits where feasible and otherwise compare to the reported figures with that caveat stated.
Why not just…?
Why not just pretrain on the synthetic data uniformly? This is the real threat. Synthetic monotonic ranking may deliver most of the gain regardless of order, and with a strong pretrained backbone the curriculum effects reported for small networks (AR-CNN) or randomly initialized models (function-class study) may shrink. We do not argue it away. The matched uniform ablation is the headline comparison, and we will report the synthetic dataset as a standalone contribution if order turns out not to matter.
Why not just pretrain on 2D quality data, as PAD-Net does? 2D pretraining never presents interocular inconsistency. Amir et al. find that 2D metrics and 2D-screen labels track headset preferences poorly. We include a 2D-pretrained variant of the same backbone so the comparison is direct rather than inferred.
Why not just oversample asymmetric examples? If asymmetric distortions merely suffer from being rare, reweighting fixes them without any ordering. The oversampling baseline settles this. The curriculum must beat it, not only beat uniform sampling.
Why not just use annotator agreement or soft labels on SCOPE? Those approaches address label noise, which is a different problem from learning rivalry. Training on the fraction of annotators preferring each image is a cheap, strong fix for noisy preferences, and we use soft targets during SCOPE fine-tuning for all variants equally rather than proposing an agreement-ordered curriculum, which soft labels would likely match.
Proposed Experiments
Pretraining runs with three or more seeds per variant. We then fine-tune and evaluate with standard splits on LIVE 3D Phase I and II, Waterloo-IVC Phases I and II, SOLID and SCOPE. Metrics are SROCC and PLCC for score databases and 2AFC accuracy, total and unanimous, for SCOPE.
There are three decisive checks. First, curriculum versus uniform pretraining on the Phase II asymmetric subset. Second, cross-database transfer: train on LIVE Phase II and test on Waterloo-IVC without fine-tuning, where there is room above ceiling. Third, sequential versus mixed curricula, measured as the drop in symmetric-subset accuracy after Stage C.
The diagnostic sensitivity test then asks whether winning variants weight the cleaner eye more heavily. On SOLID, we additionally check whether geometric-stage pretraining helps depth-quality ratings more than image-quality ratings, which is the separation Zhou and Wang predict.
References
- Predictive Auto-Encoding Network for Blind Stereoscopic Image Quality Assessment
- What Makes for a Good Stereoscopic Image?
- Compression Artifacts Reduction by a Deep Convolutional Network
- How does Multi-Task Training Affect Transformer In-Context Capabilities? Investigations with Function Classes
- Perceptual Depth Quality Assessment of Stereoscopic Omnidirectional Images
- Learning a No-Reference Quality Predictor of Stereoscopic Images by Visual Binocular Properties
- Self-Paced Learning for Latent Variable Models
- Stereoscopic Omnidirectional Image Quality Assessment Based on Predictive Coding Theory