Source-linked AI summary
AgriCountDINO: Parameter-Efficient Exemplar-Guided Counting and Localization in Agriculture
Shengjie Guo, Xin Li, Borjana Arsova, Hanno Scharr, Silvio Salvi
TL;DR
Agricultural counting and localization must handle targets that vary widely across species and imaging conditions, while exemplar boxes specify targets without category-specific retraining. AgriCountDINO uses exemplar-conditioned frozen multiscale features to predict target points, and reports lower counting errors on TPC-268 and zero-shot FSC-147 than the compared methods.
Problem
Agricultural targets vary across species, organs, growth stages, scales, and imaging conditions, while category-specific models are difficult to extend across diverse targets.
Source: §1 · p. 1Method
AgriCountDINO conditions frozen multiscale DINOv3 features on exemplar appearance and size, decodes target points, and uses recovery supervision and exemplar-adaptive duplicate suppression.
Source: §1 · p. 1Results
11.92 three-shot validation MAE on TPC-268 is 9.7% below TasselNetV4, and 14.25 zero-shot MAE on FSC-147 is 6.0% below the strongest compared method.
Source: §1 · p. 1, §3.4 · p. 4Takeaways & Limitations
AgriCountDINO combines counting with individual target locations and transfers to unseen generic object categories after training only on TPC-268.
Source: §3.4 · p. 4, §4 · p. 4Dense or visually ambiguous instances and annotation discrepancies can contribute to remaining errors; plausible predictions on unannotated targets may increase measured error.
Source: §3.3 · p. 3
Abstract
from arXiv · showhide
Accurate counting and localization of plants and their organs support phenotyping and yield estimation, yet target appearance, scale, and density vary widely across species and imaging conditions. Exemplar boxes specify the target without category-specific retraining, and point predictions identify the individual instances contributing to the count. We introduce AgriCountDINO, a parameter-efficient exemplar-guided framework for joint counting and localization. It conditions frozen multiscale DINOv3 features on exemplar appearance and size, then progressively decodes them into target points. Missed-object recovery extends supervision to targets overlooked by initial matching, and exemplar-adaptive point NMS filters duplicate predictions according to exemplar scale. With 8.4M trainable parameters, approximately one-tenth of TasselNetV4's, AgriCountDINO achieves a three-shot MAE of 11.92 on the TPC-268 benchmark, reducing counting error by 9.7\% while providing individual target locations. Trained only on TPC-268, it achieves a zero-shot MAE of 14.25 on unseen generic object categories in FSC-147, improving upon the best compared zero-shot method by 6.0\% without target-domain training or fine-tuning.
1. INTRODUCTION
AgriCountDINO addresses the need to count and localize varied agricultural targets using exemplars rather than category-specific retraining. It combines exemplar-conditioned point prediction with compact training and reports results on TPC-268 and unseen FSC-147 categories.
1. INTRODUCTION
- Agricultural targets vary across species, organs, growth stages, scales, and imaging conditions, motivating exemplars that specify targets without category-specific retraining.Counting supports yield estimation and phenotyping, while instance locations support spatial analysis and inspection.Source: §1 · p. 1
- TasselNetV4's local-count predictions do not explicitly identify individual targets, motivating spatially detailed exemplar correspondences for localization.The introduction identifies dense self-supervised DINOv3 features as a basis for learning those correspondences.Source: §1 · p. 1
- AgriCountDINO counts and localizes diverse plants and plant organs with 8.4M trainable parameters.It conditions multiscale point prediction on exemplar appearance and size, with missed-object recovery during training and duplicate suppression at inference.Source: §1 · p. 1
- 11.92 three-shot validation MAE on TPC-268 is 9.7% below TasselNetV4; zero-shot FSC-147 MAE of 14.25 is 6.0% below the strongest compared method.The FSC-147 evaluation uses no target-domain training or fine-tuning.Source: §1 · p. 1
2. METHOD
AgriCountDINO conditions frozen multiscale image features on exemplar appearance and size, then fuses them into dense point predictions. Training matches and recovers targets for localization supervision, while inference suppresses duplicate points according to exemplar scale.
2.1. Multiscale Exemplar Conditioning
- Exemplar appearance and size tokens condition frozen feature maps at multiple scales through scale-specific query modules.The conditioned maps are progressively fused from coarse to fine for high-resolution point scoring and offset prediction.Source: Figure 1 · p. 2, §2.1 · p. 2
2.2. Progressive Dense Point Decoding
- Coarse-to-fine fusion aggregates conditioned maps into a dense representation, where each position predicts a target score and point offset.Offsets can cross grid-cell boundaries to reach target centers and are clipped to the image.Source: §2.2 · p. 2
2.3. Localization-Aware Learning and Inference
- Peak matching pairs confident local maxima with annotations; missed-object recovery assigns unmatched targets distinct unused queries from nearby grid neighborhoods.Recovery searches a 5 × 5 neighborhood, then a 7 × 7 neighborhood for remaining targets, while reserving queries accepted by initial matching.Source: §2.3 · pp. 2–3
- Score losses supervise selected peaks and recovered queries, while coordinate and peak losses train recovered points with a gradual warm-in over eight epochs.Recovery terms place greater weight on targets closely spaced relative to exemplar scale.Source: §2.3 · p. 3
- At inference, exemplar-adaptive point NMS suppresses nearby candidates in descending score order using a radius based on mean exemplar width and height.The retained points form the predicted set, whose cardinality gives the count.Source: §2.3 · p. 3
3. EXPERIMENTS
The experiments evaluate AgriCountDINO on TPC-268 counting and localization, zero-shot transfer to FSC-147, and component ablations. Results include validation and test comparisons, while annotation discrepancies may contribute to measured errors.
3.2. Experimental Setup
- Evaluation uses TPC-268’s 7,000-image training split, one- and three-shot settings, image-level counting metrics, regional GAME scores, and point-level precision and recall.Zero-shot transfer is evaluated on FSC-147 without target-domain training or fine-tuning.Source: §3.2 · p. 3
3.3. Counting Accuracy and Spatial Localization
- With 8.40M trainable parameters, AgriCountDINO’s three-shot TPC-268 validation MAE of 11.92 and RMSE of 28.55 reduce TasselNetV4’s corresponding errors by 9.7% and 35.0%.It has approximately one tenth of TasselNetV4’s trainable parameters and achieves the lowest validation MAE and RMSE in both exemplar settings.Source: §3.3 · p. 3, Table 1 · p. 4
- In three-shot TPC-268 validation, GAME3 falls from TasselNetV4’s 33.33 to 25.37, a 23.9% reduction at the finest reported grid level.Predicted points achieve 0.719 precision and 0.729 recall.Source: §3.3 · p. 3, Table 2 · p. 4
- Remaining validation and test errors may reflect difficult or ambiguous targets and annotation quality; plausible predictions on unannotated targets may increase measured error.Fig. 4 shows examples of apparent discrepancies between visible targets and point annotations.Source: §3.3 · p. 3, Figure 4 · p. 4
3.4. Zero-Shot Cross-Dataset Transfer
- A zero-shot FSC-147 MAE of 14.25 improves on LOCA’s 15.16 by 6.0%, after training only on TPC-268 and without target-domain training or fine-tuning.The FSC-147 examples also show localization of individual targets in unseen categories.Source: §3.4 · p. 4, Table 3 · p. 4
3.5. Ablation Study
- Replacing DINOv3 with supervised ConvNeXt-B raises MAE from 11.92 to 17.64; removing scale-wise conditioning worsens MAE and GAME3, while recovery removal lowers recall from 0.729 to 0.692.F1 remains unchanged without recovery, making improved instance coverage its clearest contribution.Source: §3.5 · p. 4
4. CONCLUSIONS
AgriCountDINO combines parameter-efficient exemplar-guided counting and localization, with validation improvements on TPC-268 and the lowest compared zero-shot MAE on FSC-147 without target-domain fine-tuning.
4. CONCLUSIONS
- AgriCountDINO uses 8.4M trainable parameters to count and localize plants and plant organs, with lower TPC-268 validation counting and spatial errors than TasselNetV4.It also obtains the lowest zero-shot MAE among compared FSC-147 methods without target-domain fine-tuning.Source: §4 · p. 4