RECOUNT: Reference-guided Counting with Synthetic Visual Exemplars

Adriano D'Alessandro; Ali Mahdavi-Amiri; Ghassan Hamarneh

Figure 1: From a reference image to precise counts in mixed category scenes. We propose RECOUNT, which automatically turns a reference image into an unlimited supply of synthetic visual exemplars (SVEs). We use SVEs to create support sets that specify both what to count and what to reject. Applied as a lightweight filter over a frozen counter’s point proposals, RECOUNT pulls a single category out of multi-category scenes where language alone cannot. Green points are accepted, red points are rejected. Each panel pairs a query image with the target (top, green) and distractor (bottom, red) SVE galleries that drive the decision. Across these challenging scenes, RECOUNT recovers the exact count.

Figure 2: A New Approach to Inference-Time Category Specification. Existing class-agnostic counters require either per-image manual bounding-box annotations at test time ( top-left ) or text prompts too coarse to specify visual identity ( bottom-left ). In contrast, our framework ( right ) specifies a category from a single real reference image, provided for both the target and its distractors. A synthetic data engine expands these lone references into a dense, visually diverse exemplar support set. Generated once per category, the support set is reusable across any number of query images. By deferring categorization until after spatial localization, RECOUNT matches each localized proposal against this support set, isolating the target from its distractors without per-image human labeling.

Figure 3: Decoupling Dense Localization from Dense Categorization. Object counting requires solving two distinct tasks: dense localization and dense categorization. Traditional models trained on single-category benchmarks (e.g., FSC147) conflate these steps; they excel at localizing objects ( center ) but fail to separate visually similar or novel categories. Decoupling these operations treats localization as a class-agnostic geometry problem, leaving semantic boundaries to be resolved independently ( right ).

Figure 4: Fine-Grained Data Generation with Image References. Standard text-to-image models struggle to synthesize exact fine-grained details from text prompts alone. Given the prompt “Poblano peppers,” and a dense description, a text-only generation ( center ) defaults to producing generic green pepper traits. By conditioning on a single real reference image ( left ), we can design synthetic training data that capture domain-specific fine-grained visual features ( right ).

Figure 5: Automated Generation of a Fine-Grained Dense Categorization Training Dataset. Our data engine synthesizes dense training scenes by coupling real-world taxonomic relationships with generative image expansion. We initialize our categories by combining the iNaturalist taxonomy with manually seeded non-natural object classes. An LLM structurally parses this input to construct a massive, unified fine-grained taxonomy, which dictates image-fetching queries across diverse online repositories. The collected real-world reference images are filtered and then paired with LLM-generated scene templates, guiding a latent diffusion model to synthesize dense, visually complex scenes with accurate fine-grained object categories.

Figure 6: Generated Fine-Grained Training Groups. Sample outputs from our automated data engine showing three distinct semantic groups ( Asteraceae flower species, shellfish, butterfly species) containing visually similar but distinct siblings. By organizing dense, multi-object scenes into contrastive sibling subsets, this data provides the explicit training signal required for our downstream dense categorization module.

Figure 7: Contrastive Training Pipeline. We train the projection head

P_{\theta}

on synthetic dense scenes, leaving the backbone and counter frozen. (Top) From a few real references per category, our engine generates a synthetic support set for a category

c

: positive exemplars

\mathcal{G}^{+}_{c}

depicting the target, and hard-negative exemplars

\mathcal{G}^{-}_{c}

depicting its visually similar distractors. A frozen image encoder

\phi

, a frozen counter, and the projector

P_{\theta}

reduce every exemplar to a support feature. (Bottom) For a synthetic scene of

c

, the frozen counter pseudo-labels each instance with a point and

\phi

produces a feature grid; indexing the grid at these points yields per-object features, which

P_{\theta}

refines through a learned residual. (Right) A multi-negative contrastive loss

\mathcal{L}_{\mathrm{contr}}

pulls each refined per-object feature toward the positive support features and away from the negative support features. Only

P_{\theta}

is trained.

Figure 8: Point Pooling for Synthetic Visual Exemplars. Each synthetic visual exemplar (SVE) in a support gallery is reduced to a single prototype. The frozen counter proposes a point on every instance in the exemplar; we sample the backbone (

\phi

) feature grid at these points by bilinear interpolation, average the sampled features into one descriptor, and project it with the head

P_{\theta}

to obtain the exemplar’s prototype. A support gallery is the set of these prototypes, one per exemplar.

Figure 9: Crop-and-Reprocess. The pre-trained vision backbones that we utilize emit a fixed

32\times 32

patch grid, so in dense regions many instances occupy a single patch and cannot be meaningfully separated, causing the filter to miss or misjudge them ( left ). We detect each dense cluster, crop it ( middle ), and re-extract features at native resolution before re-categorizing the enclosed points ( right ), recovering instances the base grid could not resolve.

Method	Reference	Inference	$\text{A}\rightarrow\text{B}$		$\text{A}\rightarrow\text{C}$
Method	Reference	Inference	MAE $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$
PSeCo +FiGO ( 2025 )	Text	TTA	19.57	30.86	-	-
DAVE +FiGO ( 2025 )	Text	TTA	21.19	31.53	-	-
CountGD +FiGO ( 2025 )	Text	TTA	13.31	23.36	-	-
GrREC +FiGO ( 2025 )	Text	TTA	16.15	26.99	-	-
CountGD ( 2024 )	BBox	Few-Shot	-	-	46.67	70.85

Table 1: Cross-domain Transfer Performance Under Varying Inference Protocols. We evaluate cross-domain transfer from the single-category FSC-147 dataset (A) ( Ranjan et al., 2021 ) to two fine-grained, multi-category counting datasets: LookAlikes (B) ( D’Alessandro et al., 2025 ) and PairTally (C) ( Nguyen et al., 2025 ) . Notation

\text{A}\rightarrow\text{B/C}

denotes the training set (left) and test set (right). Reference modalities:

\oplus

and

\ominus

indicate the inclusion of positive and negative categories, respectively; BBox and Image represent bounding boxes from the target image and separate independent reference images; and +pseudo denotes text-selected pseudo-bounding boxes. Bold indicates the best score for each inference mode, and underline marks the overall best result. Note that LookAlikes is zero-shot only and cannot be evaluated in few-shot mode. Gray shading highlights our method.

Method	Reference	Negative Test		Mosaic test
Method	Reference	NMN $\downarrow$	PCCN $\uparrow$	CntP $\uparrow$	CntR $\uparrow$
TFPOC ( 2024 )	Text ⊕	0.75	66.04	0.69	0.85
DAVE ( 2024 )	Text ⊕	1.05	37.02	0.84	0.80
DAVE ( 2024 )	Text ${}^{\ominus}_{\oplus}$	0.08	97.62	0.84	0.80
CountGD++ ( 2026 )	Text ⊕ +pseudo	0.88	62.86	0.86	0.96
CountGD++ ( 2026 )	Text ${}^{\ominus}_{\oplus}$ +pseudo	0.07	97.99	0.90	0.96

Table 2: Prompt Adherence Results. We benchmark our method against other zero-shot methods using the PrACo dataset ( Ciampi et al., 2025 ) , which measures whether a model correctly counts the prompted category. The Negative Test prompts each image with every class and reports Normalized Mean of Negative (NMN), which is the mean of the count predicted for absent classes, normalized by the ground-truth count of the present class, and Positive Class Count Nearness (PCCN), which is the percentage of images whose positive-class count is closer to ground truth than its negative-class counts. The Mosaic Test stacks the target image with a distractor-class image and counts the target over the collage, reporting counting precision (CntP) and recall (CntR); Text = text-grounded cues (and, for CountGD++, text-synthesized exemplars); +pseudo = pseudo-exemplars detected on the query image; Image = a reference image that was not derived from the query image. Best per column in bold .

Figure 10: Qualitative results on LookAlikes . Each panel shows our filter’s per-point decisions for a target category in a scene containing co-present lookalikes. In this oracle setting, the filter is applied to ground-truth points to isolate categorization from detection. Markers encode decisions: a green circle is a target point accepted (true positive), a red circle is a target point rejected (false negative), a red

\times

is a distractor point rejected (true negative), and a green

\times

is a distractor point accepted (false positive). We report recall (R), false-positive rate (FP), and F1 at operating threshold

\delta{=}0

. Examples span 24 fine-grained categories across LookAlikes , showing the filter accepts the target while rejecting near-identical siblings.

Backbone	Head	Reference	Metrics
Backbone	Head	Reference	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$
DINOv3	dino.txt	Text	0.535	0.086	0.804
DINOv3	None	Reference Only	0.624	0.083	0.809
DINOv3	None	SVE Gallery	0.849	0.116	0.913
DINOv3	RECOUNT	SVE Gallery	0.870	0.096	0.929
DINOv2	None	Reference Only	0.542	0.065	0.765

Table 3: Ablation over architectural components and reference modalities. Evaluating the impact of vision backbone architecture, the RECOUNT head, and reference modality on performance. Evaluated using the Lookalikes dataset with oracle localizations.

Number of SVEs	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$
1	0.818	0.097	0.914
3	0.833	0.090	0.914
6	0.847	0.083	0.918
12	0.870	0.096	0.929
24	0.878	0.096	0.924

Table 4: Impact of Synthetic Visual Exemplar (SVE) Gallery Size on Performance. The number of SVEs used in the positive and negative galleries are variable, and so we ablate over the number of SVEs used and it’s impact on performance. We find performance saturates when the gallery size is restricted to 12 SVEs per-category. Evaluated using the Lookalikes dataset with oracle localizations.

Operation	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$
Whole-Image (GAP, all patches)	0.746	0.123	0.890
Per-Instance Pseudo-Detections	0.825	0.098	0.916
Pooled Pseudo-Detection	0.870	0.096	0.929

Table 5: SVE Pooling Operation Ablation. SVE galleries consist of synthetic images whose visual features originate from the underlying RECOUNT module. Utilizing an SVE image as part of gallery for specifying a category necessitates a mechanism to pool its feature grid. We compare global average patch pooling against point-feature pooling via pseudo-detections, where point coordinates are extracted per image using an off-the-shelf counter (CountGD). These coordinates are indexed into the visual feature patch grid to extract interpolated point features. Evaluated using the Lookalikes dataset with oracle localizations.

Visual Exemplar	Crowding	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$
Synth.	$\times$	0.735	0.096	0.882
Synth.	✓	0.870	0.098	0.929
Real	✓	0.925	0.101	0.930

Table 6: Evaluating Exemplar Source and Quality. We evaluate the performance gap between real and synthetic visual exemplars while holding all other parameters fixed; only the per-category exemplar gallery varies. To isolate the synthetic-to-real domain gap, we also compare synthetic exemplars containing a single canonical object (no crowding) against synthetic exemplars specifically prompted to contain crowds. Evaluated using the Lookalikes dataset with oracle localizations.

Reference Used
Train	Infer	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$
$\times$	$\times$	0.727	0.121	0.887
✓	$\times$	0.807	0.127	0.903
✓	✓	0.870	0.096	0.929

Table 7: Effect of Reference Usage During Training and Inference. Ablation evaluating performance across different combinations of train- and test-time references to assess the impact of reference guidance. Evaluated using the Lookalikes dataset with oracle localizations.

Strategy	Recall $\uparrow$	FPR $\downarrow$	AUC $\uparrow$	Recall-Contam. $\uparrow$	FPR-Contam. $\downarrow$
Base	0.870	0.096	0.929	0.620	0.329
Crop	0.883	0.099	0.949	0.753	0.368
$\Delta$	+0.013	+0.003	+0.020	+0.133	+0.039

Table 8: Benefits of Crop-and-Reprocess. We evaluate the performance of our model when dense regions are adaptively cropped and then passed back through the model. We speculate that contamination from neighboring points impacts the separability of a target point. To evaluate this, we also report contaminated recall, which takes the original 32×32 grid from DINOv3, and calculates recall for all target points that have a distractor sibling within a 3×3 grid around that point. Evaluated using the Lookalikes dataset with oracle localizations.

$d$	Neighborhood	$N$	Recall $\uparrow$
$d$	Neighborhood	$N$	Base	Crop	$\Delta$
0	Same patch ( $1\times 1$ )	636	0.5466	0.733	$+0.169$
1	$3\times 3$ shell	4,055	0.717	0.818	$+0.101$
2	$5\times 5$ shell	3,905	0.844	0.886	$+0.042$
3	$7\times 7$ shell	3,000	0.891	0.914	$+0.023$
4	$9\times 9$ shell	2,445	0.909	0.928	$+0.019$

Table 9: Spatial contamination. We evaluate recall as a function of Chebyshev distance

d

(in grid cells) to the nearest distractor object. We show that dense categorization breaks down when distractor objects share a patch with the target object. However, there also appears to be a degradation in performance when adjacent distractor objects multiple patches away. Evaluated using the Lookalikes dataset with oracle localizations.

RECOUNT: Reference-guided Counting with Synthetic Visual Exemplars

Abstract

Figures & tables

Object Counting

Count Anything

CountEx: Fine-Grained Counting via Exemplars and Exclusion

AdaCount: Training-Free Similarity-Guided Spatial and Feature Adaptation for Zero-Shot Object Counting

Abstract

Object Counting

Explore similar work

Count Anything

CountEx: Fine-Grained Counting via Exemplars and Exclusion

AdaCount: Training-Free Similarity-Guided Spatial and Feature Adaptation for Zero-Shot Object Counting