Drone imagery is cheap to collect and expensive to label well. A single flight produces thousands of nearly identical overlapping frames. Annotators working on dense canopies, herds, or crowded streets routinely miss tiny objects, and the rare classes that matter most (a bicycle among cars, an uncommon tree species) make up a sliver of the boxes. The current answer from the data-curation literature is to score every training example with large foundation models, such as open-vocabulary detectors and vision-language models, and keep only the best fraction. Two problems make that answer a poor fit for drones. Those models are known to fail at localizing small aerial objects. Their forward passes over millions of high-resolution tiles can cost more compute than training the lightweight detector that will actually fly.
We propose a different source of quality judgment. Train a small committee of different lightweight detectors on disjoint folds of the training set. Each member scores the images it did not see, and the members' agreement and disagreement with the human labels decide what to keep, what to flag as mislabeled, and what to drop as redundant. The committee exists only during curation. What ships is a single detector trained on the curated set, so the deployed model carries none of the usual cost of running an ensemble at test time. The claim is that this in-domain committee finds aerial label errors and redundancy more reliably than foundation-model scoring, at a small fraction of its cost. We judge it by the total budget, curation plus training, not by training compute alone.
Core Claim
For small-object aerial detection, quality signals should come from models that can actually see the objects. A committee of out-of-fold, architecturally diverse small detectors can do three jobs. It finds missing boxes when the members confidently agree on an object that the annotation lacks. It finds wrong class labels when the members agree with each other but not with the label. It separates redundant images from informative ones: redundant images are those every member already gets right and that duplicate other frames, while informative images are those where the members disagree even though the labels look correct. We predict that a detector trained on a committee-curated subset of roughly 30–50% of the images will match or beat the same detector trained on all images. The gain should be largest on rare classes, and the curated detector should come out ahead at matched total compute. We also predict that foundation-model scoring will cost an order of magnitude more and will be worse at flagging small-object label errors.
Where This Sits in Prior Work
The closest work is Quality over Quantity: Boosting Data Efficiency Through Ensembled Multimodal Data Curation. EcoDatum combines heavy operators (GroundingDINO for local box-text alignment, Flip-CLIP for global alignment, perceptual-hash deduplication) through a Snorkel-style label model. It kept 3.5M of 12.8M DataComp pairs and placed first on the small track. It supplies our aggregation idea, combining noisy votes into a single quality score. Its authors flag the cost of the operators as an open problem, and its target is web image-text pairs rather than box-annotated nadir imagery. We replace its general-purpose operators with in-domain detectors and score individual boxes rather than captions.
SmokeBench constrains the design. It reports that general multimodal language models reach near-zero box IoU on smoke, Grounding DINO reaches 0.245, and a dedicated YOLOv8n reaches 0.773. That is direct evidence that the operators EcoDatum relies on are unreliable judges of where small aerial things are. It is also the reason we expect a cheap in-domain committee to judge boxes better.
The text-curation line provides the selection machinery we borrow from. D4 deduplicates semantically and keeps cluster prototypes. SHED scores clusters by approximate Shapley contribution. Bhatt et al. select a diverse covering set in one pass and argue that iterative active learning is too costly. All three keep "typical" data and discard outliers, which is risky when the rare classes are the outliers. We therefore add an explicit per-class retention floor.
Deep learning‐based ecological analysis of camera trap images motivates the long-tail concern. It shows that species-level estimates collapse under label noise and data reduction even when coarse metrics survive. It used only random pruning and uniform noise, which is the gap we fill with algorithmic curation and a structured noise protocol.
On the ensemble side, VisDrone-DET2019 shows that ensembles dominate drone detection: the DPNet-ensemble reached 29.62% AP against 16.09% for Cascade R-CNN. They pay for this at inference. Fei et al. show the same pattern for stacked regressors on UAV multispectral yield data. We move the ensemble's error-cancelling benefit from inference time to curation time.
The language-model cost line enters as the foil and as the accounting lens. Routers such as ECVL-ROUTER save inference cost by sending easy queries to small models. We apply the same instinct to curation: use the cheapest judge that is competent for the task.
Research Program
The pipeline has four stages, and every arm of the study shares the first one. In the first stage, perceptual-hash deduplication removes near-identical frames, so that later gains cannot be credited to trivial redundancy removal.
In the second stage, the training images are split into $K=4$ folds. Each committee member is trained on three folds on a shortened schedule and run on the fourth. Members differ in architecture so that their errors are less correlated: a YOLO-family detector with a shallow high-resolution head in the style of LSNET, a lightweight-neck variant in the style of Tassel-YOLO, and a Swin-backbone detector.
In the third stage, the out-of-fold predictions are matched to annotations box by box. Each image receives three scores: a missing-label score, a class-conflict score, and an informativeness score, which is committee disagreement on boxes that agree with the label. These signals are combined with a weak-supervision label model, reusing EcoDatum's aggregation.
In the fourth stage, images are selected to fill a target budget under a per-class floor. Images containing rare-class boxes are never dropped unless they are flagged as noisy. Flagged images are either removed or, in a second variant, relabeled with the committee's consensus boxes.
All costs are reported in GPU-hours and charged against the method. The quantity of interest is total compute,
$$C_{\text{total}} = C_{\text{curate}} + R \cdot C_{\text{train}}(S),$$
where $C_{\text{curate}}$ is the cost of building and running the committee (or the foundation-model operators), $C_{\text{train}}(S)$ is the cost of one training run on the selected subset $S$, and $R$ is the number of training runs that reuse the curated set. We report results at $R=1$, the harshest case, and at $R=10$, which corresponds to a modest hyperparameter sweep.
Proposed Experiments
We use three benchmarks from the file. VisDrone-DET2019 provides the long tail: bicycles reach about 13% AP against about 51% for cars in the challenge results. SelvaBox (paper, resource) provides dense tropical crowns, where we expect incomplete annotation. DroneVehicle provides an RGB-infrared vehicle detection setting. If the raw video frames of the Maize Tassel UAV Dataset are obtainable, it would add a heavy-redundancy setting; the file reports 40,000 raw frames.
Because the true noise rate of public labels is unknown, we also inject structured noise into clean splits at 10%, 20%, and 40% of boxes, following the noise levels of the camera-trap study. The noise takes the forms drones actually produce: deleted boxes, swapped classes, and boxes shifted off their objects. The injected noise is known, which lets us measure the precision and recall of noise flagging directly. The metrics are [email protected] and AP, per-class AP with the rarest three classes reported separately, noise-flag precision and recall, and $C_{\text{total}}$. Every arm uses the same deployed detector, schedule, and hyperparameter budget.
Baselines
The naive reference point is the deployed detector trained on all deduplicated images, alongside random subsets at the same budgets as the curated subsets. We first reproduce full-data numbers close to published ones as a sanity check. The file reports LPANet at 77.7% mAP on DroneVehicle (Wu et al.) and the DPNet-ensemble at 29.62% AP on VisDrone. These are architecture results, not curation results, so we compare against our own full-data run of the same detector rather than against these figures. Beating the naive reference means equal or higher overall mAP and higher rare-class AP at no more than half the images, with total compute at $R=1$ no greater than full-data training.
The strongest existing curation method is EcoDatum-style scoring, adapted to boxes. GroundingDINO alignment between each annotated box and its class name, plus an open vision-language model asked to verify image-level labels, are combined with the same label model, the same per-class floor, and the same budgets. The file reports EcoDatum's 0.182 average on the DataComp small track. No reported result was found on Lacuna for EcoDatum-style curation of UAV detection data, so we run it ourselves. The committee beats it if it gives higher downstream rare-class AP and higher noise-flag recall at lower $C_{\text{curate}}$.
We also include D4-style prototype selection on detector embeddings, without a class floor, to test whether keeping typical data harms the tail.
Three ablations isolate the key ingredient. The first replaces the committee with a single out-of-fold model of the same total compute. The second replaces it with a homogeneous committee: the same architecture with different seeds. The third removes the per-class floor. If the heterogeneous committee does not beat the single model, the ensemble is not earning its place.
Why not just…?
Why not just filter by a single model's training loss? Dropping high-loss examples is a standard cheap noise filter, and it is the real threat to this proposal. A model scoring its own training data partly memorizes the noise. A single held-out model also cannot tell "hard but correct" from "mislabeled," because both produce high loss. Committee agreement against the label is what separates the two cases. The single-out-of-fold-model ablation, matched in compute, settles this. If it ties the committee, we will report that the ensemble adds nothing and recommend the single model.
Why not just deduplicate with perceptual hashes? For redundancy alone this may be most of the gain, which is why every arm already includes it. Hashing cannot detect missing boxes or swapped classes, and the camera-trap results suggest those errors are what destroy rare-class estimates. Our claim therefore rests on the gain on top of deduplication. If that gain is negligible on real (uninjected) labels, the honest conclusion is that public drone datasets are clean enough that hashing suffices.
Why not just use foundation models, as EcoDatum does? SmokeBench's 0.245 IoU for Grounding DINO against 0.773 for a tiny dedicated detector suggests these models misjudge small aerial boxes. Their cost per high-resolution tile is also large. The head-to-head arm measures both the cost and the noise-flag recall, rather than assuming either.
Why not just ensemble at deployment, or train a bigger detector? Test-time ensembles multiply latency and power by the number of members, which UAV hardware cannot absorb. The camera-trap study found that architecture depth had minimal effect on ecological metrics, and SoybeanNet's largest variant was not its best. We include a test-time committee as a reference upper bound, so readers can see how much of its accuracy the single curated detector recovers.
Decisive Checks
Three outcomes decide the proposal. First, on VisDrone with 20% injected noise, a committee-curated 40% subset should beat full-data training on bicycle and tricycle AP, and also beat the foundation-model arm, at lower $C_{\text{curate}}$. Second, on SelvaBox, the images the committee flags for missing crowns should, when re-annotated by hand on a sample of 200 images, show a measurably higher missing-crown rate than random images. Third, the heterogeneous committee should beat the compute-matched single model. If the first and third checks fail, the idea reduces to deduplication plus loss filtering, and we would say so.
References
- Quality over Quantity: Boosting Data Efficiency Through Ensembled Multimodal Data Curation
- SmokeBench: Evaluating Multimodal Large Language Models for Wildfire Smoke Detection
- D4: Improving LLM Pretraining via Document De-Duplication and Diversification
- SHED: Shapley-Based Automated Dataset Refinement for Instruction Fine-Tuning
- An Experimental Design Framework for Label-Efficient Supervised Finetuning of Large Language Models
- Deep learning‐based ecological analysis of camera trap images is impacted by training data quality and quantity
- VisDrone-DET2019: The Vision Meets Drone Object Detection in Image Challenge Results
- Assessment of Ensemble Learning to Predict Wheat Grain Yield Based on UAV-Multispectral Reflectance
- ECVL-ROUTER: Scenario-Aware Routing for Vision-Language Models
- An Efficient Algorithm for Small Livestock Object Detection in Unmanned Aerial Vehicle Imagery
- Tassel-YOLO: A New High-Precision and Real-Time Method for Maize Tassel Detection and Counting Based on UAV Aerial Images
- SelvaBox: A high‑resolution dataset for tropical tree crown detection
- Large Language Model Guided Progressive Feature Alignment for Multimodal UAV Object Detection