Semi-supervised object detection trains a detector from a few labeled images plus many unlabeled ones. A "teacher" copy of the model labels the unlabeled images, and its confident guesses, called pseudo-labels, become training targets. On aerial and surveillance imagery this breaks down for tiny objects. The teacher sees a downsampled image, so a 10-pixel vehicle becomes a few blurry feature cells and receives low confidence. A fixed confidence cutoff then throws these objects away, and the model learns to ignore small things. The accepted fix is to lower the cutoff for small objects. That recovers them, but it also admits more wrong labels.
We propose a different fix. When the teacher is unsure about a small candidate, look again at full resolution before deciding. The deployed "student" must stay small and fast. The teacher only runs during training, so we can afford to give it native-resolution crops of the few regions where its decision is ambiguous. We cap that extra work with a budget measured in wall-clock milliseconds on real hardware, not in parameter or operation counts. In two sentences: tiny objects look uncertain mainly because they were downsampled, not because they are intrinsically hard, so the right response is to restore resolution at the point of decision, not to relax the decision. We test whether a small, hardware-measured budget of high-resolution teacher re-scoring yields cleaner small-object pseudo-labels than size-aware threshold lowering, and whether that cleaner signal survives distillation into a latency-optimal low-resolution student.
Core Claim
The closest existing work is Rethinking Scale Imbalance in Semi-supervised Object Detection for Aerial Images (S³OD). It documents a "confidence gap": small objects get systematically lower teacher confidence on DOTA-v1.5. S³OD responds with size-aware percentile thresholds, Wasserstein-distance label assignment, and a rule that stops unlabeled small candidates from being treated as background. That work motivates ours, because it established the gap. We part ways with it on the remedy. S³OD treats the gap as a property of the score distribution and moves the cutoff. We treat it as a symptom of lost input resolution and change the input the teacher scores. The thresholds stay strict.
Our claim makes a falsifiable prediction that S³OD's view does not. If the gap is caused by resolution, re-scoring true small objects at native resolution should close much of it. Re-scoring false small candidates the same way should not raise their confidence. Under that condition, verification gives higher pseudo-label precision at equal recall than threshold lowering.
The second ingredient comes from the analysis of uncertain regions in Combating Noise: Semi-supervised Learning by Region Uncertainty Quantification. That paper finds mislabeling concentrated in a "danger zone" near decision boundaries. It handles those regions by down-weighting them or pushing them toward background. Our reading is that this danger zone is exactly where extra evidence is worth buying. We re-examine those regions rather than discount them.
Research Program
The teacher first runs as usual on the downsampled unlabeled image. Each candidate box gets a confidence $p$ and an apparent size. Candidates above a strict upper cutoff are accepted, and those below a lower cutoff are rejected. Both cutoffs follow an exponential moving average of past confidences, as in the relative-threshold schemes of Taejong Joo et al., so they track the teacher as it matures.
Candidates between the cutoffs form the ambiguous band. Each ambiguous candidate $i$ gets a priority $a_i = p_i(1-p_i)/\sqrt{A_i}$, where $A_i$ is its box area in pixels. The first factor is the ambiguity of a binary keep/discard decision. The second favors small boxes, where downsampling destroys the most information. Each candidate also has a measured re-scoring cost $c_i$: the wall-clock time of one teacher pass on its native-resolution crop, profiled on the training GPU. The crop sizes are fixed to a few power-of-two shapes so this cost is stable.
We choose which candidates to verify by solving
$$\max_{S} \sum_{i \in S} a_i \quad \text{subject to} \quad \sum_{i \in S} c_i \le B,$$
where $S$ is the set of candidates sent for verification and $B$ is a per-image latency budget in milliseconds. We solve it greedily by ratio $a_i/c_i$. Overlapping candidates share one crop, much like the adaptive search window of Global-Local MAV Detection under Challenging Conditions based on Appearance and Motion (GLAD). The difference is that GLAD's local crop serves inference, is triggered by motion, and fails on hovering targets. Ours serves label verification at training time and is triggered by confidence, so a static target does not defeat it.
Verified candidates are kept only if the native-resolution score clears the strict upper cutoff. Candidates that are not verified are neither kept as foreground nor counted as background. They are marked "ignore", following the trust-mask idea of Auto-Annotation Quality Prediction for Semi-Supervised Learning with Ensembles but without its 18-model ensemble. We also re-run box regression on the crop. This addresses the localization noise isolated by Pseudo-label Correction and Learning For Semi-Supervised Object Detection (PCL). PCL refines boxes by repeatedly jittering them on the same low-resolution features. We instead provide new pixel evidence once.
The student is trained at deployment resolution against these verified labels, so it learns to predict from degraded inputs what the teacher confirmed from clean ones.
Parameter-latency decoupling sets the constraints in two places. First, the budget $B$ is in measured milliseconds, because operation counts mispredict real cost. DeepRebirth: Accelerating Deep Neural Network Execution on Mobile Devices and the AIM 2019 constrained super-resolution findings in Zhang et al. both show that fragmented or misaligned operations dominate runtime. Second, the student is chosen by measured latency on an NVIDIA Jetson NX Xavier, the edge board GLAD reports on. Following the depth-width findings of Nemotron-Flash, we compare a parameter-minimal deep-narrow student against a wider, power-of-two-aligned student of equal measured latency.
This comparison raises a question the existing literature leaves open: does a latency-matched but larger student absorb cross-resolution supervision better? If it does, the parameter count of a "lightweight" student is the wrong axis even for training dynamics, not only for speed.
Baselines
The naive reference is a supervised-only detector trained on the labeled split at deployment resolution. Alongside it we run a plain mean-teacher detector with a fixed 0.9 cutoff, Unbiased Teacher for Semi-Supervised Object Detection. That method updates the teacher as a slow moving average of the student and uses focal loss. It is the standard pipeline that every newer method modifies, so our verification is a drop-in change to it.
The strongest competitor on this exact question is S³OD. It targets the same failure mode (small objects lost to strict thresholds on aerial images) on the same benchmark. It reports 40.02 mAP on DOTA-v1.5 with 1% labeled data, a +10.72 gain over its supervised baseline. We will reproduce S³OD within our codebase with the same detector, labeled splits, schedule, and hyperparameter search budget, and report both our reproduction and the published number. Beating it means higher overall mAP and higher small-object AP at 1% and 5% labels, at an equal deployed student latency on the Jetson. We also require total training wall-clock within a stated factor of S³OD's, which we set at 1.3×.
Because S³OD's pieces are complementary to ours, we also run S³OD with verification added. Without that run, a gain could come from bolting extra machinery onto S³OD rather than from verification itself.
PCL is the strongest pseudo-label-correction competitor. It reports 36.47 mAP on MS COCO with 10% labels, +2.43 over SoftTeacher. We run it on DOTA-v1.5 under the same budget. For PCL on DOTA, as with GLAD-style detectors in a semi-supervised setting, we found no reported result on Lacuna, so we will measure it ourselves.
Three ablations isolate the key ingredient. In the first, matched-compute random verification, the budget is spent on randomly chosen ambiguous candidates rather than by priority. This tests whether selection matters. In the second, same-resolution re-scoring, the same candidates are re-scored on upsampled low-resolution crops rather than native-resolution crops. This tests whether new pixels matter or merely a second look. In the third, a FLOP budget replaces the millisecond budget, which tests whether measured latency changes which candidates get verified.
Why not just…?
Why not just lower the threshold for small objects, as S³OD does? This is the main threat, and S³OD's results show it works well. Our argument is that lowering the cutoff trades precision for recall, while verification can raise both. The cheapest experiment in the program decides this. On DOTA-v1.5 unlabeled images, where ground truth exists, we re-score the teacher's small candidates at native resolution and record two things: how much true positives' confidence rises, and whether false positives' confidence stays flat. If true and false candidates rise together, resolution does not separate them, and the method reduces to threshold lowering at extra cost. We would then drop verification and keep only the latency-matched student analysis.
Why not just run the teacher at full resolution on tiles over the whole image? Full tiling is the natural upper bound on pseudo-label quality, and it may simply win. We report it explicitly as an oracle-like baseline. The question is cost. In mean-teacher training the teacher labels every unlabeled image at every step, so full tiling multiplies teacher cost by roughly the number of tiles. Our claim is that a budget of 10–20% of full-tiling teacher time recovers most of its small-object AP gain. We plot small-object AP against teacher milliseconds for both. If the budgeted curve does not dominate the cheap end of the tiling curve, selectivity has no value.
Why not just train and deploy the student at higher resolution, or add content-aware pooling? The deployment constraint rules out the first. The latency of multi-scale pipelines on edge hardware is the very problem being avoided. Content-aware downsampling is a real, complementary option. LIP: Local Importance-Based Pooling raises supervised Faster R-CNN small-object AP on COCO from 21.9% to 24.0%. However, LIP's pooling module adds non-standard operations, and the latency literature above warns these can cost more wall-clock time than their parameter count suggests. We therefore include a LIP student, report its measured Jetson latency, and test whether verified pseudo-labels and LIP gains stack.
Why not just use a bigger teacher or an ensemble? A larger teacher still sees the same downsampled pixels, so its errors on tiny objects should be largely shared with the student. The ensemble of Simon et al. needs 18 forward passes per image. We compare against a 2× wider teacher, and against a 3-model ensemble at matched teacher milliseconds, to test whether capacity or resolution is the binding factor.
Proposed Experiments
The main benchmark is DOTA-v1.5 at 1%, 5%, and 10% labeled splits. Over 57% of targets in the DOTA family are 10–50 pixels, per DOTA: A Large-Scale Dataset for Object Detection in Aerial Images. We report mAP, AP by size bin, and, using the held-out ground truth of the unlabeled pool, pseudo-label precision and recall by size bin over training. That last measurement makes the noise mechanism visible rather than inferred.
As a transfer check where targets are extremely small, we use ARD-MAV, whose objects occupy about 0.02% of the frame. GLAD's fully supervised 0.80 AP there is the reference ceiling for a semi-supervised student at reduced resolution. COCO at 10% labels, reporting AP_s, checks that the method does no harm outside aerial imagery.
All deployed students are timed on a Jetson NX Xavier at batch size 1, with the median of repeated runs. Teacher costs are timed on the training GPU.
Three results would decide the proposal. The first is the native-resolution re-scoring separation test described above, which comes before anything else. The second is whether native-resolution verification beats same-resolution re-scoring at equal budget; this confirms that new pixel evidence, not a second look, carries the effect. The third is whether, at equal measured Jetson latency, the wide aligned student gains more small-object AP from verified labels than the parameter-minimal student. That result would link parameter-latency decoupling to how well a model learns from pseudo-labels, not just how fast it runs.
References
- Rethinking Scale Imbalance in Semi-supervised Object Detection for Aerial Images
- Combating Noise: Semi-supervised Learning by Region Uncertainty Quantification
- Improving self-training under distribution shifts via anchored confidence with theoretical guarantees
- Global-Local MAV Detection Under Challenging Conditions Based on Appearance and Motion
- Auto-Annotation Quality Prediction for Semi-Supervised Learning with Ensembles
- Pseudo-label Correction and Learning For Semi-Supervised Object Detection
- DeepRebirth: Accelerating Deep Neural Network Execution on Mobile Devices
- AIM 2019 Challenge on Constrained Super-Resolution: Methods and Results
- Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models
- Unbiased Teacher for Semi-Supervised Object Detection
- LIP: Local Importance-Based Pooling
- DOTA: A Large-Scale Dataset for Object Detection in Aerial Images