Small-Target-Aware Label Assignment (STAL)

• The paper presents novel pseudo-box, surrogate box, and Gaussian receptive-field methods that decouple candidate selection from true box geometry to ensure positive assignments for tiny objects.
• STAL is a label assignment strategy that improves detection recall by guaranteeing at least one positive candidate per ground truth, boosting AP for sub-stride objects.
• Empirical results demonstrate improved AP on datasets like AI-TOD and COCO, highlighting STAL's practical impact on dense and aerial object detection.

Small-Target-Aware Label Assignment (STAL) refers to a family of label assignment strategies in dense object detection designed to guarantee positive sample assignment for extremely small objects that would otherwise fail to trigger anchor- or point-based assignment mechanisms. Standard assignment paradigms frequently yield zero or near-zero positive assignments for sub-stride objects, leading to severe recall degradation in tiny or small-object detection. STAL decouples candidate selection from true box geometry, incorporates receptive-field or pseudo-box relaxation, and optionally incorporates Gaussian receptive field matching to ensure every ground truth, regardless of size, produces positive training signal.

1. Motivation and Problem Statement

Conventional label assignment protocols—anchor-based (e.g., IoU-threshold, box-overlap) or anchor-free (e.g., center/point-in-box)—are inherently biased against small targets. When the size of a ground-truth bounding box is less than the feature-map stride, no anchor or feature-map center may fall within the box after quantization or downsampling. This results in zero candidates marked positive and eliminates gradient flow for such targets. This “scale–sample imbalance” induces large gaps in recall and mean average precision (AP) for small and tiny targets, particularly evident in aerial, SIRST, and dense indoor benchmarks (Xu et al., 2022, Dai et al., 2022, Jocher et al., 2 Jun 2026, Guan et al., 3 Jan 2026).

STAL addresses these limitations by modifying the geometric filter used in candidate selection, relaxing the strict spatial requirements that suppress small or sub-pixel-aligned boxes. The core aim is to guarantee that every true object, irrespective of its original size, receives sufficient (≥1) positive assignments per feature level, enabling detectors to propagate learning signals from all objects present in the ground truth.

2. Core Methodologies

STAL encompasses several methodological branches, each converging on the same objective: robust, scale-invariant label assignment. The canonical strategies demonstrated in the literature are as follows.

A. Pseudo-Box Clamping (All-Scale Pseudo-Box Assignment)

• For each GT box $(c_x, c_y, h, w)$ and feature level with stride $s$, construct a pseudo-box $B_p$:

$$B_p = \begin{cases} (c_x, c_y, h, w), & h \cdot w > p^2 \ (c_x, c_y, p, p), & \text{otherwise} \end{cases}$$

with $p = \alpha s$ ($\alpha \in [1.0, 2.0]$). Positive assignment occurs if a candidate location’s coordinates fall within $B_p$; this strategy is widely used in SIRST and OSCAR frameworks (Dai et al., 2022).

B. Surrogate Box Assignment (Clamped Candidate Mask)

• For each ground truth $g_i = (x_i, y_i, w_i, h_i)$, construct a surrogate box $\tilde g_i = (x_i, y_i, \tilde w_i, \tilde h_i)$ with dimensions

$$\tilde w_i = \begin{cases} s_{\text{ref}} & w_i < s_{\min} \ w_i & \text{otherwise} \end{cases}, \quad \tilde h_i = \begin{cases} s_{\text{ref}} & h_i < s_{\min} \ h_i & \text{otherwise} \end{cases}$$

where $s$0 is the minimal stride and $s$1 is a chosen reference stride. Candidate selection uses $s$2 for mask formation, but regression uses the original $s$3 (Jocher et al., 2 Jun 2026).

C. Gaussian Receptive-Field Modeling

• Represent both each GT and each feature location as 2D Gaussians, with mean as center and variances as either half the theoretical receptive field (for features) or half the box size (for GTs). Similarity is then measured either by Kullback–Leibler divergence, Wasserstein-2 distance, or Gaussian Combined Distance, yielding a Receptive-Field Distance (RFD) or normalized similarity (Xu et al., 2022, Guan et al., 3 Jan 2026).

D. Hierarchical or Adaptive Positive Supplementation

• Combine point-prior initialization with an RFD-based ranking. After marking initial positives (e.g., center region or pseudo-box), select supplementary positives among the remaining locations using a ranked RFD. Ambiguous matching selects candidates with intermediate RFD scores for inclusion (Guan et al., 3 Jan 2026).

3. Development of STAL Variants

Several prominent detectors and assignment modules implement STAL mechanisms, adjusted to their architectural and task-specific constraints.

Method	Principle	Key Technical Distinction
OSCAR (SIRST)	All-Scale Pseudo-Box	Pseudo-box expansion, all-level assign
YOLO26	Surrogate Box Assignment	Size clamping in candidate mask
RFLA	Gaussian Receptive-Field Modeling	Receptive-field distance, hierarchy
RFAssigner	Mixed Point + RF Assignment	Adaptive positive supplementation

• OSCAR (Dai et al., 2022) employs all-scale pseudo-box expansion so that every GT activates at every FPN stage and at least one location per level is marked positive regardless of true GT size. The pseudo-box side $s$4 is tuned for optimal balance between recall and false positives.
• YOLO26 (Jocher et al., 2 Jun 2026) integrates a candidate mask clamping step for Task-Aligned Learning (TAL), so that surrogate boxes ensure every object—including those smaller than the smallest stride—receives positive assignment without altering downstream regression or scoring.
• RFLA (Xu et al., 2022) and RFAssigner (Guan et al., 3 Jan 2026) leverage Gaussian receptive-field models to translate every GT-feature pair into a continuous affinity, facilitating robust assignments even in the absence of hard spatial overlap.

4. Algorithmic Implementation and Hyperparameterization

The implementation of STAL typically requires minimal modification to the overall training pipeline but introduces key hyperparameters:

• Pseudo-box side: $s$5, $s$6 tuned in the range $s$7; optimal values near $s$8.
• Surrogate/reference stride: $s$9, typically set as the second-smallest stride for maximal coverage without over-clustering positives.
• Top-k hierarchical or scale-wise selection: Number of locations (per GT per scale) assigned as positives; e.g., $B_p$0 in RFLA, $B_p$1 for scale-wise supplementation in RFAssigner.
• Supplement thresholds: Statistical thresholding using mean $B_p$2 of RFD among top candidates.
• RFD windows for ambiguous matching: E.g., $B_p$3, $B_p$4 in RFAssigner (Guan et al., 3 Jan 2026).

No changes are made to inference code; all modifications occur in training-time label assignment. Anchor and point priors may be entirely removed or retained depending on the framework.

5. Quantitative Impact and Empirical Analysis

STAL mechanisms consistently improve both overall AP and, critically, AP on small objects ($B_p$5) across datasets characterized by numerous tiny targets, including AI-TOD, VisDrone2019, DOTA-v2.0, TinyPerson, and MS-COCO.

• On AI-TOD, (Xu et al., 2022) reports Faster R-CNN: 11.1 → 21.1 AP, DetectoRS: 20.8 → 24.8 AP with RFLA. RFAssigner (Guan et al., 3 Jan 2026) achieves up to 22.3 AP (AP_vt 7.5, AP_t 22.2, AP_s 27.1).
• For OSCAR on SIRST-V2, pseudo-box side $B_p$6 yields maximum mNoCoAP of 77.6%, compared to 71.9% with center-only assignment (Dai et al., 2022).
• YOLO26 (Jocher et al., 2 Jun 2026) achieves $B_p$729.6 AP$B_p$8 on COCO val2017 with STAL enabled, compared to 29.0 without.
• Ablations confirm optimal performance with moderate pseudo-box expansion and moderate supplementary positive supplementation, avoiding the decrease in precision incurred by excessive expansion.

6. Connections to Related Assignment Strategies

STAL generalizes and unifies multiple recent approaches:

• Receptive-Field Label Assignment (RFLA) and RFAssigner (Xu et al., 2022, Guan et al., 3 Jan 2026): Use ERF-based Gaussian modeling and probabilistic similarity metrics to achieve continuous, scale-invariant assignments.
• Task-Aligned Learning (TAL): Standard TAL scoring frameworks can incorporate STAL as a pre-filtering or masking step without disrupting regression, classification, or matching loss definitions (Jocher et al., 2 Jun 2026).
• Center/Point priors: STAL’s adaptive supplementation or pseudo-box expansion can completely replace or robustly complement fixed-ratio center sampling.

The central conceptual advance is the decoupling of candidate filter geometry from the true box size, either by explicit pseudo-box or surrogate-box enlargement or by shifting from hard binary to continuous affinity assignment.

7. Practical Implications and Limitations

STAL-style schemes yield substantial improvements in detection recall and localization precision for tiny objects, have virtually zero inference overhead, and integrate seamlessly into both anchor-based and anchor-free pipelines. The only notable training cost is minor increased memory or computation for per-level candidate masks or affinity matrices. Over-expansion of the candidate region can reduce precision, so hyperparameters must balance recall and false positive rates.

A plausible implication is that dense detection frameworks seeking robust performance in remote sensing, traffic surveillance, or microscopy—where small and tiny object recall dominates—will treat STAL as an essential assignment component. There are no reported negative consequences for medium or large object AP, and scale-invariance is preserved or enhanced.

Key references: (Xu et al., 2022, Dai et al., 2022, Jocher et al., 2 Jun 2026, Guan et al., 3 Jan 2026).