Center-DETR: Center-Aware DETR for Cytology

• The paper introduces a center-aware detection reformulation that predicts cell centers rather than regressing variable box sizes.
• It employs a Co-DINO framework with a Swin-Large backbone and center-preserving crop augmentation to align supervision with fixed-size annotations.
• The approach achieves competitive performance in cervical cytology challenges by optimizing loss tuning and applying analytical box reconstruction.

Center-DETR denotes a center-aware reformulation of a Co-DINO / Co-DETR detector for cervical cytology, introduced for the RIVA Cervical Cytology Challenge and built around the observation that the dataset’s annotations are rigid fixed-size bounding boxes. In that setting, the principal learning problem is not free-form box extent estimation but accurate localization of cell centers and, for Track A, Bethesda-category classification. The method therefore retains a strong DETR-family pipeline—specifically Co-DINO with a Swin-Large backbone—but reorganizes target definition, augmentation, post-processing, and loss emphasis around center prediction. Under this formulation, the reported system achieved 1st place in Track B and 2nd place in Track A (Kong et al., 2 Apr 2026).

1. Problem definition and representational shift

Center-DETR is motivated by a dataset protocol in which cell instances are annotated with boxes described as rigidly fixed at $100 \times 100$ pixels. That annotation regime changes the semantics of detection. In ordinary object detection, regressing width and height is meaningful because object extent varies and the box is intended to capture object geometry. In the RIVA setting, the box size is largely annotation-imposed rather than object-specific, so the critical variable is the cell center rather than a freely varying $(w,h)$ pair (Kong et al., 2 Apr 2026).

The method therefore models detection as prediction of object centers together with class probabilities. For Track A, the class space comprises 8 Bethesda categories plus 1 background class; for Track B, localization is the primary objective. The central rationale is that standard IoU-driven box regression can become counterproductive under fixed-size annotations: a detector with a slightly shifted center may compensate by predicting larger-than-needed boxes to recover overlap, which introduces size-regression noise and degrades mAP (Kong et al., 2 Apr 2026).

This formulation makes Center-DETR best understood not as a wholly new DETR-family architecture from scratch, but as a task-specific adaptation of an existing end-to-end detector to a center-point prediction problem. The resulting detector is intended to answer where each cell center is, and, when required, which Bethesda category it belongs to.

2. Architectural basis in Co-DINO and Co-DETR

The underlying detection framework is Co-DINO within the broader Co-DETR collaborative hybrid assignment training paradigm. The backbone is Swin-Large, chosen in preference to ViT-Large because the target domain contains tiny cells, dense local structures, and subtle spatial morphology; the paper attributes to Swin hierarchical multi-scale representations with localized attention and progressively built hierarchical semantics that are better suited to such imagery (Kong et al., 2 Apr 2026).

The overall pipeline consists of an input image, a Swin-Large backbone for hierarchical multi-scale feature extraction, and a Co-DINO detector with object queries and denoising queries during training. In addition to the main DETR-style branch, the training graph includes one-to-many auxiliary branches: Faster R-CNN, ATSS, RetinaNet, and FCOS. These heads provide dense supervision during training and are discarded at inference time (Kong et al., 2 Apr 2026).

Among the auxiliary branches, FCOS is especially aligned with the center-point formulation. That alignment is not used to redefine the main DETR decoding mechanism, but it reinforces the center-aware training signal inside the collaborative hybrid assignment framework. Inference uses the main Co-DINO-Swin detector, with decoder outputs interpreted as center coordinates and class scores rather than unconstrained object boxes (Kong et al., 2 Apr 2026).

The architecture is therefore hybrid in a specific sense: the principal prediction pathway remains one-to-one DETR-style set prediction, while auxiliary one-to-many detectors enrich supervision during training. Center-DETR does not discard DETR matching or decoder-based prediction; it changes what the detector is expected to predict.

3. Center-aware augmentation and supervision

A central customization is center-preserving crop augmentation. Standard random cropping in natural-image detection can retain any object whose box overlaps the crop, but under the cytology formulation this can preserve truncated cells whose centers fall outside the visible region. The method instead keeps an object after cropping if and only if its center lies inside the crop region: $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$ Objects whose centers fall outside the crop are discarded (Kong et al., 2 Apr 2026).

The significance of this rule is domain-specific. The paper argues that Pap-smear interpretation depends on morphological integrity, including nucleus appearance and nucleus-to-cytoplasm ratio. If a crop removes the nucleus center, the remaining fragment may cease to be a valid positive example. Center-preserving cropping therefore aligns augmentation with the center-based target definition and reduces supervision noise from partial cells (Kong et al., 2 Apr 2026).

Its standalone quantitative effect is modest but consistent. On Track B, Co-DINO-Swin improves from $0.609$ to $0.611$, and Co-DINO-ViT improves from $0.604$ to $0.606$, when center-preserving cropping is added alone (Kong et al., 2 Apr 2026). This suggests that the crop policy is not the dominant source of performance gain, but it regularizes training in a direction consistent with the task formulation.

4. Analytical geometric box optimization

The most distinctive component is an analytical post-processing rule that reconstructs a fixed-size square from a predicted center. Because ground-truth boxes are $100 \times 100$, the paper asks a specific question: given center localization jitter, what fixed square side length $S$ should be used at inference to maximize expected IoU with the ground truth? The reported answer is

$S = 101.5.$

The final box is therefore reconstructed from the predicted center using a fixed side length of $(w,h)$0 (Kong et al., 2 Apr 2026).

The derivation is based on overlap geometry. Let $(w,h)$1 and $(w,h)$2 be center offsets between the predicted and ground-truth boxes. The overlap width and height are defined as

$(w,h)$3

$(w,h)$4

With intersection area $(w,h)$5 and union $(w,h)$6, the paper optimizes the expected IoU over the empirical localization jitter distribution $(w,h)$7 (Kong et al., 2 Apr 2026).

The logic is a tradeoff. If $(w,h)$8, any center error immediately reduces overlap. If $(w,h)$9, there is a tolerance margin

$$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$0

so small center errors can be absorbed without immediate intersection loss. But if $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$1 becomes too large, the union area grows and IoU falls. Under typical jitter of about $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$2 pixels, the optimum reported in the paper is $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$3 (Kong et al., 2 Apr 2026).

This geometric prior is not merely a post hoc adjustment for the proposed detector. The paper reports model-agnostic gains on Track B when the same $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$4 reconstruction rule is applied to several baselines, indicating that the benefit is tightly linked to the fixed-size annotation regime rather than to Swin or Co-DINO alone (Kong et al., 2 Apr 2026).

Detector	Track B before $$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$5 after box optimization
YOLO	$$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$6
RetinaNet	$$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$7
CenterNet	$$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$8
Co-Deformable-DETR	$$x_{\min} \le c_{x_i} \le x_{\max}, \qquad y_{\min} \le c_{y_i} \le y_{\max}.$$9
Co-DINO-ViT	$0.609$0

A plausible implication is that, when annotation geometry is largely artificial, analytically matching that geometry can be more important than learning width and height as free semantic variables.

5. Optimization, assignment structure, and empirical performance

Center-DETR inherits the Co-DETR training structure: a main one-to-one DETR branch and auxiliary one-to-many heads. The total loss is given as

$0.609$1

Here $0.609$2 is the classification loss on the main branch, $0.609$3 is the box regression loss, $0.609$4 is generalized IoU loss, and $0.609$5 aggregates auxiliary-head losses over one-to-many assignments (Kong et al., 2 Apr 2026).

A central design choice is to downweight IoU-based supervision because the center-plus-fixed-size reconstruction prior reduces the need for network-driven size regression. The paper sets

$0.609$6

Track-specific reweighting then emphasizes the dominant difficulty of each challenge track. For Track A, where the principal bottleneck is classification among similar Bethesda categories, the paper uses

$0.609$7

For Track B, where localization is the objective, it uses

$0.609$8

Within the center-based formulation, the Track B choice effectively emphasizes center precision (Kong et al., 2 Apr 2026).

Implementation details reported in the paper include PyTorch and MMDetection, initialization from COCO pre-trained weights, training on the official training split with validation-set evaluation, and final challenge submission after merging training and validation data. Training used an NVIDIA RTX 4090 with 48 GB VRAM, and inference is reported at 12 GB VRAM. Aside from the explicit loss reweighting, the authors state that default hyperparameters were used (Kong et al., 2 Apr 2026).

The main comparative results show the strength of the underlying Co-DINO-Swin detector and the incremental value of the center-aware modifications. Reported baselines are: YOLO, Track A $0.609$9, Track B $0.611$0; RetinaNet, $0.611$1 and $0.611$2; CenterNet, $0.611$3 and $0.611$4; Co-Deformable-DETR, $0.611$5 and $0.611$6; Co-DINO-ViT, $0.611$7 and $0.611$8; and Co-DINO-Swin, $0.611$9 and $0.604$0 (Kong et al., 2 Apr 2026). On Track B, the Co-DINO-Swin ablation proceeds as follows:

Configuration	Track B
Baseline	$0.604$1
$0.604$2 Center-preserving crop	$0.604$3
$0.604$4 Crop $0.604$5 box optimization	$0.604$6
$0.604$7 Crop $0.604$8 box optimization $0.604$9 loss tuning	$0.606$0

The ablation shows that the largest gain comes from analytical box optimization, not from cropping or loss tuning alone. This supports the paper’s claim that the decisive issue is the match between prediction geometry and the fixed-size annotation regime (Kong et al., 2 Apr 2026).

6. Relation to adjacent center-based DETR research, assumptions, and scope

Within arXiv literature, Center-DETR should be distinguished from several adjacent but non-identical uses of centers in detection and retrieval. In overhead imagery, centerpoint-only supervision has been studied through adaptations of RetinaNet and Faster R-CNN that predict object centers rather than boxes, showing that centerpoint labels can preserve most center-localization performance on xView, DOTA 1.5, and FAIR1M; however, those models are anchor-based CNN detectors rather than DETR-style set predictors (Inder et al., 2022). In video moment retrieval, DETR-based models already predict temporal spans in $0.606$1 form, and length-aware matching has been proposed to improve short-moment center prediction indirectly through length-specialized queries rather than through a new center head (Park et al., 2024).

At the level of DETR query semantics, SAP-DETR offers an explicit counterpoint to center-prior designs. It argues that centralizing query reference points can reduce query saliency and create redundant predictions, and proposes query-specific salient points with $0.606$2 box parameterization instead of center-based semantics (Liu et al., 2022). In dense pedestrian detection, PED analyzes why vanilla DETR and deformable DETR underperform in crowd scenes and responds with dense queries, rectified attention fields, visible-part supervision, and faster matching; that work is query-based and box-guided rather than center-supervised, but it is relevant to any Center-DETR-like design intended for dense scenes (Lin et al., 2020).

Against that background, the most precise characterization of Center-DETR is narrow and task-specific. It is a center-aware Co-DINO / Co-DETR detector specialized for a fixed-size annotation regime in cervical cytology, not a universal redefinition of DETR around center-only prediction (Kong et al., 2 Apr 2026). Its assumptions are explicit: fixed-size annotations must be meaningful, center localization must carry most of the information needed for detection, localization jitter should be small enough that $0.606$3 remains near-optimal, and center-preserving cropping should be appropriate for the domain (Kong et al., 2 Apr 2026).

The paper also states that the method is largely engineering-driven to maximize mAP, with less emphasis on deeper clinical or biological modeling. This suggests that the design is highly effective for challenge conditions yet somewhat specialized. A plausible implication is that the transferable insight is conditional rather than universal: if a dataset’s annotation protocol makes box size largely artificial, then a DETR-family detector can profitably be reformulated around center prediction and an analytically chosen geometric prior; if object size and shape are semantically meaningful, the same reformulation may become suboptimal (Kong et al., 2 Apr 2026).