Show It What to Count, Then Make It Point

TL;DR for operators

Agricultural counting systems face a practical problem when the target changes. A workflow that counted tassels yesterday may need to count fruit, flowers, or another plant structure tomorrow, under different scales and densities. Retraining a dedicated model for every such variation makes the counting system itself part of the operational bottleneck.

AgriCountDINO¹ tests a different arrangement: show the system one or a few example boxes indicating what should be counted, keep its pretrained visual encoder frozen, and train a comparatively small set of components to match those examples across the image and return explicit target points. In three-shot TPC-268 validation, it reports MAE 11.92 versus 13.20 for TasselNetV4, with 8.40M versus 84.67M trainable parameters. The distinction matters: 8.40M is not the total size of the DINOv3-backed model.

For an operator, the interesting output is therefore not only a lower benchmark error. Each predicted object also has a location that can be inspected or passed into spatial analysis. The external-transfer result is encouraging but bounded: training on TPC-268 and evaluating on FSC-147 without target-domain training or fine-tuning yields MAE 14.25, but this is one external dataset and AgriCountDINO does not lead every reported metric.

The target can change without redefining the whole model

Many agricultural counting problems combine two sources of variation. The scene changes—camera conditions, density, growth stage—and the definition of the target changes as well. A fixed-category model absorbs that definition during training. Changing the target can therefore require another training cycle or another model.

AgriCountDINO instead receives exemplar boxes specifying the object of interest in the current image. The task becomes conditional visual matching: find other image locations that resemble the supplied examples closely enough to count as the same target.

That change in task definition is central to the architecture. A DINOv3-pretrained ConvNeXt-B encoder extracts visual features at four scales, but its parameters remain frozen. The trainable system around it extracts exemplar appearance at each scale, combines that information with exemplar width and height, and uses cross-attention to condition image features on what the operator has shown it.

The resulting features are then fused from coarse to fine. Coarser representations contribute semantic information about what resembles the exemplar; progressively finer maps restore spatial detail. A dense prediction head assigns target confidence and a two-dimensional point to candidate locations.

Parameter efficiency here is therefore an adaptation property, not evidence that the entire vision system contains only 8.40M parameters. The paper reports the number of trainable parameters; it does not report 8.40M as the total size of the frozen-backbone system.

A count becomes more useful when the system can account for its objects

On three-shot TPC-268 validation, AgriCountDINO records MAE 11.92, RMSE 28.55, and $R^2$ of 0.93. TasselNetV4 reports 13.20, 43.93, and 0.83 respectively. The paper also evaluates how well counts remain correct within increasingly fine image subdivisions. At GAME3, AgriCountDINO reaches 25.37 versus 33.33 for TasselNetV4.

That spatial result deserves separate attention from image-level MAE. A model can produce a plausible total while distributing errors across the image. AgriCountDINO instead produces explicit predicted points and reports three-shot point precision of 0.719 and recall of 0.729.

Those points are not bounding boxes or segmentation masks. They answer a narrower question: which locations contributed individual instances to this count? For phenotyping, review, or spatial aggregation, that can make an aggregate prediction easier to audit.

Crowded scenes create a specific training problem. An annotated target may fail to generate a distinct initial score peak, leaving it without a matched prediction. The paper’s missed-object recovery procedure searches nearby unused dense queries and assigns additional supervision to such unmatched annotations.

Its ablation suggests that recovery mainly affects coverage. Removing it raises MAE from 11.92 to 12.71 and reduces recall from 0.729 to 0.692, while reported F1 remains 0.723.

The ablations separate representation, conditioning, and recovery

The component tests are best read as architectural diagnostics rather than additional headline benchmarks.

The largest MAE deterioration occurs when the DINOv3 representation is replaced. That is consistent with the paper’s mechanism: reusable dense visual correspondences in the frozen representation provide much of the substrate on which exemplar matching operates.

Scale-wise conditioning then determines how that substrate is redirected toward the current target. Missed-object recovery addresses a more specific failure mode after the basic correspondence machinery is already working.

Zero-shot transfer still uses examples at inference

The FSC-147 experiment is easy to overread. “Zero-shot” here means the model receives no FSC-147 target-domain training or fine-tuning. It does not mean the counting process runs without exemplar guidance.

After training on TPC-268, AgriCountDINO reports FSC-147 MAE of 14.25, compared with 15.16 for LOCA, 17.88 for CACViT, and 19.17 for TasselNetV4. That result supports the possibility that the learned counting machinery can operate beyond the agricultural categories used for training.

But MAE does not tell the whole comparison. AgriCountDINO’s zero-shot RMSE is 106.88, while CACViT reports 82.57. The external-transfer evidence also comes from one dataset. The experiment is therefore evidence for cross-dataset reuse, not evidence of universal category transfer.

The in-domain results require similar discipline. AgriCountDINO is strong on the highlighted three-shot validation metrics, but it does not dominate every test metric or every comparator. For example, LOCA reports a lower three-shot TPC-268 test MAE, 17.51 versus 21.76 for AgriCountDINO.

The business case is reusable targeting plus inspectable output

Cognaptus inference: for agricultural vision teams, the architecture suggests a different unit of reuse. Instead of treating each species or plant organ as a reason to build another counting model, a workflow could preserve a common visual representation and redirect the counting behavior through exemplars.

That could affect two operational decisions. First, teams evaluating new target categories can ask whether examples are sufficient to redirect an existing system before committing to another category-specific training pipeline. Second, downstream users can inspect predicted target locations rather than receiving only a scalar count.

The parameter result also changes what “efficient adaptation” means. Only a relatively small task-specific component is trained, but the frozen encoder still has to exist and run. The paper therefore supports reduced trainable parameter requirements, not a claim that deployment compute or total model memory falls in the same proportion.

The boundary is dense scenes, annotation quality, and limited external evidence

The paper identifies dense or visually ambiguous targets as remaining difficult. It also notes apparent annotation discrepancies in TPC-268: a model can predict a visually plausible target that is not annotated and still be penalized by the benchmark.

Those two issues affect different decisions. Crowding concerns the model’s ability to separate nearby instances. Annotation discrepancies concern whether the benchmark itself fully represents the objects a human reviewer might regard as legitimate targets.

Neither invalidates the reported comparison, but both limit how directly benchmark error should be translated into field reliability.

AgriCountDINO’s more durable contribution is therefore architectural. It shows that agricultural counting can be organized around a frozen reusable representation, target examples supplied at inference, and explicit instance points rather than category-specific aggregate counting alone. The benchmark results make that design credible enough to test in broader deployments. They do not yet establish how broadly it will transfer once the crops, sensors, annotation rules, and operating environments change.

Cognaptus: Automate the Present, Incubate the Future.