Teaching Low-Resolution Students Only What They Can See: Resolution-Aware Targets for Fine-Grained Semi-Supervised Learning

Fine-grained recognition means telling apart classes that look almost identical, such as two sparrow species or two variants of one airliner. It gets much harder when the camera delivers only a few dozen pixels per object, because the distinguishing details (a bill stripe, an engine cowling) may cover one or two pixels. Labels for such data are also expensive, since only experts can provide them. A natural fix is to borrow the asymmetric "teacher sees an easy view, student sees a hard view" recipe from semi-supervised learning, letting resolution play the part of the hard view. A slowly updated teacher looks at a sharp, high-resolution image of an unlabeled example and produces a soft label. The student sees the same image shrunk to 56×56 or 32×32 and is trained to match that label.

That recipe has a hidden flaw, and it is the subject of this proposal. The high-resolution teacher often knows things that cannot be recovered from the low-resolution image, and these are exactly the fine distinctions between confusable sibling classes. Forcing the student to reproduce them asks it to guess, which is a noisier version of the confirmation bias that already undermines pseudo-labeling on fine-grained data. We propose splitting the teacher's target into two parts. The first is a coarse part, which group of look-alike classes the image belongs to, and the student always matches it. The second is a fine part, which class within that group, and the student matches it only in proportion to how recoverable that group's distinctions are at the student's resolution. Recoverability is measured online from the student's own confusion on labeled low-resolution images.

Core Claim

Cross-resolution consistency on unlabeled data should help low-resolution fine-grained classifiers. The naive version, however, wastes capacity and corrupts decisions within confusable groups. A target that respects what the student can actually resolve should beat both naive matching and generic target smoothing, and its gains should concentrate on confusable class groups.

The objective follows from the chain rule for the Kullback–Leibler divergence (KL, the standard measure of mismatch between two probability distributions). Let $p_T$ and $p_S$ be the teacher's high-resolution and the student's low-resolution class posteriors. Let $G_1,\dots,G_K$ be a partition of the classes into confusion groups, $P(k)=\sum_{c\in G_k}p(c)$ the group probability, and $p(c\mid k)$ the within-group distribution. The loss is

$$ \mathcal{L} = \mathrm{KL}\big(P_T \,\|\, P_S\big) + \sum_{k=1}^{K} w_k\, P_T(k)\, \mathrm{KL}\big(p_T(\cdot\mid k)\,\|\,p_S(\cdot\mid k)\big). $$

Here $w_k\in[0,1]$ is the recoverability of group $k$ at the student's resolution. Setting every $w_k=1$ gives back plain full-distribution matching exactly. The proposal therefore changes a single quantity: how much weight the within-group term receives. We set $w_k$ to the student's chance-corrected within-group accuracy on a held-out labeled low-resolution split, clipped to $[0,1]$ and smoothed over epochs. The groups come from spectral clustering of the student's symmetrized low-resolution confusion matrix and are recomputed each epoch. This grouping technique is taken directly from HD-CNN, which clusters a confusion matrix to route confusable classes to specialist sub-classifiers. We use the clusters to shape targets, not to build an architecture.

Relation to Prior Work

Several papers motivate the gap. Fine-Grained Adversarial Semi-Supervised Learning uses covariance pooling and adversarial entropy optimization to do semi-supervised fine-grained classification. It shows that FixMatch-style hard thresholding reinforces errors among confusable classes, but it works only at full resolution. A2-SSL learns fine-grained hash codes without labels. It shows that destructive augmentations erase fine cues, which is why its positive views are only resized or cropped. That finding constrains our design: our student receives no color jitter, blur or erasing, and resolution is its only strong perturbation. A2-SSL never varies resolution and studies retrieval rather than classification.

Other papers occupy the resolution side and supply our strongest comparisons. ViMD distills a high-resolution Vision Mamba teacher into a low-resolution student that uses a frozen super-resolution front end, with full labels throughout. RACNN trains super-resolution layers end-to-end with the classification loss, and Singh et al. regularize low-resolution features with hand-annotated attributes. All three are fully supervised, and none asks which teacher distinctions a low-resolution student can actually learn.

AugKD is the closest work mixing asymmetry with resolution. It uses unpaired zoom augmentations and invertible transforms for teacher–student super-resolution distillation, but its targets are pixels, not classes. The teacher–student machinery itself comes from Unbiased Teacher, whose teacher is an exponential moving average (EMA, a running average of the student's weights), and from Soft Teacher, which weights pseudo-labels by confidence instead of thresholding them.

Research Program

One backbone, a ResNet-50 with resolution-specific batch normalization, serves both views. The EMA teacher processes a weakly augmented high-resolution crop. The student processes the same crop downsampled with antialiasing to a resolution drawn from {32, 56}, then upsampled to the network's input size, matching the protocol of prior low-resolution work. Labeled images train both resolutions with cross-entropy. Unlabeled images contribute the loss above, weighted by the teacher's confidence rather than filtered by a hard threshold. Only the low-resolution student is deployed. This mirrors the realistic setting where high-resolution training images exist but the deployed sensor is low-resolution, the same assumption ViMD makes.

The analysis has three threads. First, an identifiability diagnostic. On a split held out from fully labeled data, we compare a supervised low-resolution model with the high-resolution teacher. We measure how often the teacher's within-group decision is something the low-resolution model gets wrong on that same image. This tells us whether naive matching transmits a large or a negligible amount of unlearnable signal.

Second, we test whether the learned groups track true visual confusability. CUB is the natural test case, because hierarchical structure there is meaningful. We check whether $w_k$ rises smoothly from 32×32 to 56×56.

Third, as a secondary arm, we replace the student's strided downsampling with the learnable importance-weighted pooling of LIP. LIP raised small-object detection AP on COCO from 21.9% to 24.0% but has never been tested on fine-grained classes. The question is whether it raises $w_k$, meaning it makes more fine distinctions recoverable rather than only improving accuracy overall.

Baselines

The naive reference point is a ResNet-50 trained with cross-entropy on the labeled low-resolution images alone. It is the floor any use of unlabeled data must clear. Singh et al. report that this kind of model reaches 38.09% at 32×32 on AwA-2, against 86.79% at 224×224, which shows how severe the degradation is. We report the same model with all labels as a ceiling reference.

FixMatch on low-resolution images, with its standard RandAugment strong view and hard threshold, is the off-the-shelf semi-supervised competitor. Mugnai et al. report that it reaches 57.4% on iNaturalist-Aves at full resolution, versus 65.4% for their method. We also run Mugnai et al.'s adversarial semi-supervised method at low resolution, which is the strongest existing semi-supervised fine-grained method in the file. No reported result for either method at low resolution was found on Lacuna.

Naive cross-resolution consistency is our loss with every $w_k=1$. It is the obvious idea at this intersection and the comparison the proposal stands or falls on. A frozen high-resolution teacher, trained on the labeled split and then distilled into the low-resolution student over labeled plus unlabeled images, is the ViMD-style alternative brought into the semi-supervised setting.

The strongest supervised low-resolution reference is ViMD, which reports 80.19% top-1 on CUB-200-2011 at 56×56 with full labels. That figure is 2.35 points above the earlier SRKD distillation method. RACNN reports 51.3% to 58.1% on CUB and 50.4% to 63.8% on Stanford Cars at 50×50 with AlexNet. Both are cited as references with full labels, not as matched competitors, because they use different backbones and more supervision.

Two ablations isolate the key ingredient. In the first, matched-entropy temperature smoothing, the teacher's full distribution is softened with one temperature per image, chosen so its entropy equals that of our structured target. This separates "less confident targets" from "targets shaped by confusion structure." In the second, random groups keep the group sizes but scramble which classes belong together.

For fairness, every semi-supervised method uses the same backbone, labeled and unlabeled splits, image budget, and training length, plus a fixed hyperparameter search of equal size on the held-out labeled split. Beating the baselines means two things. Our method must exceed naive cross-resolution consistency and the frozen-teacher variant in top-1 accuracy on at least two of three datasets at each resolution, beyond the spread across three seeds. It must also exceed temperature smoothing, with most of the gain appearing in within-group accuracy.

Why not just…?

Why not just super-resolve the image with a pretrained model and run a normal fine-grained classifier? RACNN directly compares a staged pipeline, with super-resolution trained for reconstruction, against end-to-end training, and the staged pipeline loses on every dataset. Super-resolution trained to reconstruct pixels invents texture that carries no class information. We nonetheless include a pretrained super-resolution front end, combined with the full-label classifier, as an extra row, so readers can see the gap at our resolutions.

Why not just run FixMatch on the low-resolution images? Mugnai et al. show that FixMatch's hard threshold reinforces wrong pseudo-labels among confusable classes, an 8-point deficit on iNaturalist-Aves. A2-SSL shows that strong photometric augmentation erases fine cues. At 32×32 little such detail remains to erase. FixMatch is still run as a baseline, so this claim is tested, not assumed.

Why not just distill from a frozen high-resolution teacher trained on the labels? This is a real threat, because it is simple and uses the same unlabeled data. Our loss is orthogonal to the choice of teacher. We therefore run the structured target with both an EMA teacher and a frozen teacher. If the frozen teacher with naive matching equals our method, the contribution collapses to "use unlabeled data for cross-resolution distillation," and we will report it that way.

Why not just soften the teacher's targets with a temperature or label smoothing? If the problem is overconfident targets, uniform softening might be enough. The matched-entropy ablation settles this. A structured target concentrates its softening inside groups the student cannot resolve, while uniform smoothing also blurs the coarse decisions the student can make reliably. If the two perform the same, confusion structure adds nothing and the core claim fails.

Proposed Experiments

The benchmarks are CUB-200-2011, Stanford Cars and FGVC-Aircraft. Each is downsampled to 56×56, following ViMD, and to 32×32, following Singh et al. We use 10% and 30% labeled splits, with the remaining training images unlabeled. A small labeled hold-out is used to estimate $w_k$ and is never used for testing. Metrics are top-1 accuracy, within-group accuracy (accuracy given that the predicted group is correct), group accuracy, and calibration error on within-group decisions.

The decisive checks are three. First, the identifiability diagnostic must show a non-trivial fraction of teacher within-group decisions that cannot be recovered at low resolution. If that fraction is small, the method has no target, and we expect it to tie naive matching. Second, our method must beat naive matching and matched-entropy smoothing mainly through within-group accuracy and calibration. Third, the advantage should be larger at 32×32 than at 56×56, because fewer fine distinctions survive at the lower resolution. If the advantage does not grow, the recoverability weighting is not doing what we claim.