Receptive-Field Label Assignment (RFLA)

• Receptive-Field Label Assignment (RFLA) is a strategy using Gaussian modeling to assign labels based on receptive-field distances, targeting improved tiny-object detection.
• It replaces traditional heuristics by measuring sample-to-ground-truth affinity, ensuring more balanced positive sample generation.
• Variants like RFAssigner incorporate adaptive soft assignment and multi-scale receptive field models to enhance detection performance across diverse object sizes.

Receptive-Field Label Assignment (RFLA) is a label-assignment strategy for object detection that replaces conventional box-prior or point-prior heuristics with a Gaussian receptive-field prior, and measures sample-to-ground-truth affinity by a receptive-field distance. In the original formulation, RFLA was introduced for tiny object detection through Gaussian modeling of the Effective Receptive Field (ERF), a Receptive Field Distance (RFD), and a Hierarchical Label Assignment (HLA) module (Xu et al., 2022). Subsequent work realized the same receptive-field perspective in a generic dense-detector setting through RFAssigner, which combines a point prior with adaptive Gaussian Receptive Field (GRF) supplementation inside a soft-assignment framework (Guan et al., 3 Jan 2026).

1. Problem setting and motivation

RFLA addresses a specific weakness of standard dense detection pipelines: the generation of too few positive samples for small or tiny objects. In anchor-based detectors such as Faster R-CNN and RetinaNet, and in anchor-free single-stage detectors such as FCOS and AutoAssign, training samples are divided into positive and negative sets by spatial heuristics such as IoU thresholds, point-inside-box rules, or centerness. Small ground-truth boxes have few feature-map locations within their extents and low IoU with anchors, so only a handful of positives are generated for tiny objects. This scale imbalance starves small-object detectors of positive training signals and is associated with weak small-object performance (Guan et al., 3 Jan 2026).

The original RFLA paper states that either box prior in the anchor-based detector or point prior in the anchor-free detector is sub-optimal for tiny objects, and attributes this to many outlier tiny-sized ground-truth samples under current label-assignment paradigms. Its central claim is that IoU-threshold based and center-sampling strategies are skewed to large objects, and that a receptive-field-based criterion can produce more balanced learning for tiny objects (Xu et al., 2022).

Later work distinguishes between hard and soft assignment. Modern methods such as AutoAssign and DW use soft assignment, assigning each location a positive weight and a negative weight, which enables end-to-end learning of where positives should lie. However, because these methods still rely on point priors or IoU priors, they retain the small-object scarcity problem. The same later account characterizes the original RFLA as an anchor-like hard assignment that ranks locations by receptive-field matching and selects a hierarchical set of positives; this can boost tiny-object recall but may conflict with soft assignment and over-emphasize tiny targets at the expense of medium and large objects (Guan et al., 3 Jan 2026).

2. Gaussian receptive-field modeling and receptive-field distance

The mathematical core of RFLA is the Gaussian modeling of both detector receptive fields and ground-truth boxes. In the original formulation, the ERF of each feature-map location $(x_n,y_n)$ is modeled as a 2D Gaussian

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$

with $\sigma_x=\sigma_y=er_n$, where the radius $er_n$ is half the theoretical receptive field $tr_n$ of layer $n$:

$$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$

The corresponding ground-truth box $(x_g,y_g,w_g,h_g)$ is also represented as a Gaussian,

$$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$

The ERF Gaussian is written as $N_e(\mu_e,\Sigma_e)$ with $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$0 and $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$1 (Xu et al., 2022).

The original paper evaluates two closed-form distances between these Gaussians. The first is the squared 2-Wasserstein distance,

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$2

The second is the Kullback–Leibler divergence,

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$3

In either case, the paper normalizes the distance into a zero-one RFD score by

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$4

This construction makes the receptive-field criterion directly comparable across candidate locations (Xu et al., 2022).

RFAssigner preserves the Gaussian view but changes both the receptive-field model and the distance. It models each ground-truth box as

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$5

and each feature point’s theoretical receptive field at location $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$6 as

$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$7

where $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$8 are the full theoretical RF diameters on the FPN level, and the implementation uses four scales: $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$9, $\sigma_x=\sigma_y=er_n$0, $\sigma_x=\sigma_y=er_n$1, and $\sigma_x=\sigma_y=er_n$2. To compare the two Gaussians, RFAssigner uses the symmetric, scale-invariant Gaussian Combined Distance (GCD),

$\sigma_x=\sigma_y=er_n$3

and then defines the receptive-field distance as

$\sigma_x=\sigma_y=er_n$4

In the reported ablation, this GCD-based RFD gives the best cross-scale behavior among the tested alternatives on AI-TOD-v2 (Guan et al., 3 Jan 2026).

3. Assignment algorithms: HLA and adaptive supplementation

The original RFLA introduces Hierarchical Label Assignment (HLA), which begins by computing a matrix of RFD scores,

$\sigma_x=\sigma_y=er_n$5

It then applies a two-stage ranking procedure. In Stage 1, for each ground truth $\sigma_x=\sigma_y=er_n$6, the algorithm picks the top $\sigma_x=\sigma_y=er_n$7 feature indices $\sigma_x=\sigma_y=er_n$8 maximizing $\sigma_x=\sigma_y=er_n$9 and sets $er_n$0. In Stage 2, each $er_n$1 is shrunk by a factor $er_n$2, the scores are recomputed as $er_n$3, and for each ground truth not yet covered the algorithm picks the single feature with maximal $er_n$4 and sets $er_n$5. All other assignments remain zero. The reported settings are $er_n$6 and $er_n$7, and the intended effect is that every ground truth gets at least one positive sample while high-score points can receive up to $er_n$8 assignments (Xu et al., 2022).

RFAssigner retains a two-stage organization but changes the logic of both stages. Stage 1 follows FCOS or DW: initial positives are all feature-map locations whose projected center lies inside a ground-truth box and whose predicted scale level is appropriate. This produces the point mask $er_n$9. Stage 2 then supplements positives from the set of unassigned candidates. For each ground truth, all unassigned locations are scored by RFD at the four TRF scales; the method takes the top $tr_n$0 candidates by RFD, computes their mean $tr_n$1 and standard deviation $tr_n$2, and supplements the positive mask with any candidate whose $tr_n$3 (Guan et al., 3 Jan 2026).

The same stage includes ambiguous matching. Candidate locations whose max-RFD lies in $tr_n$4 are allowed to be considered, are ranked by RFD, and are added subject to the same $tr_n$5 criterion into a supplementary mask $tr_n$6. The final positive mask is

$tr_n$7

This design is explicitly intended to add high-quality but missing positives while focusing on hard but ambiguous samples across objects. A plausible implication is that RFAssigner preserves the coverage guarantees of point priors while using receptive-field similarity to repair their scale bias (Guan et al., 3 Jan 2026).

4. From hard assignment to dual-weight soft assignment

RFLA was introduced as a replacement for IoU- or center-sampling-based assigners in both one-stage and two-stage detectors. The implementation summary states that it slots neatly into two-stage and one-stage detectors, with reported settings including SGD for 12 epochs, learning-rate decay at epochs 8 and 11, RPN proposals of 3000 for two-stage backbones, inference confidence threshold $tr_n$8, NMS IoU $tr_n$9, and keeping the top 3000 boxes. For FCOS, centerness is modified with a small constant $n$0 to avoid gradient explosions (Xu et al., 2022).

RFAssigner is explicitly framed as a soft-assignment strategy and follows DW’s Dual-Weight paradigm once $n$1 has been fixed. Each location $n$2 receives a learned positive weight $n$3, and each location $n$4 but within a looser centerness prior receives a negative weight $n$5. The classification and regression losses are

$n$6

$n$7

The backbone configuration is ResNet-50 + FPN using levels P3–P6, while RFAssigner* also uses P2. No changes are made to the feature extractor or detection head; the method only intercepts the target generation stage to compute per-location positive and negative masks and weights. The formulation is described as fully end-to-end differentiable and incurring no inference-time overhead (Guan et al., 3 Jan 2026).

The contrast between the two formulations is methodologically important. The original RFLA uses a receptive-field criterion as the assignment mechanism itself, whereas RFAssigner embeds receptive-field reasoning inside a soft-weighting framework. This suggests a shift from purely geometric hard selection toward a hybrid of geometric prior and learned weighting.

5. Empirical behavior across datasets

The original RFLA reports substantial gains on tiny-object benchmarks. On AI-TOD with Faster R-CNN+FPN, the baseline IoU assigner yields $n$8, adding RFD alone yields $n$9, and adding RFD with full HLA yields $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$0. Under full HLA, the choice of distance gives $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$1 with GIoU using an ERF-sized box prior, and $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$2 with either $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$3 or $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$4. On AI-TOD with a ResNet-50 backbone, Faster R-CNN + RFLA reaches $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$5 AP, Cascade R-CNN + RFLA reaches $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$6 AP, and DetectoRS + RFLA reaches $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$7 AP. The same paper also reports gains on TinyPerson, VisDrone, and DOTA-v2.0, including FCOS on TinyPerson from $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$8 to $$tr_n = tr_{n-1} +(k_n-1)\prod_{i=1}^{n-1}s_i, \qquad er_n=\tfrac12\,tr_n.$$9, Faster R-CNN on VisDrone from $(x_g,y_g,w_g,h_g)$0 to $(x_g,y_g,w_g,h_g)$1, and DetectoRS on DOTA-v2.0 from $(x_g,y_g,w_g,h_g)$2 to $(x_g,y_g,w_g,h_g)$3 (Xu et al., 2022).

RFAssigner evaluates three datasets with distinct object scale distributions: AI-TOD-v2, VisDrone-2019, and MS-COCO-2017. AI-TOD-v2 contains 28 036 aerial images and 700 621 instances with mean object size approximately $(x_g,y_g,w_g,h_g)$4 px and uses metrics including AP@[.50:.05:.95], AP$(x_g,y_g,w_g,h_g)$5, AP$(x_g,y_g,w_g,h_g)$6(2–8 px), AP$(x_g,y_g,w_g,h_g)$7(8–16 px), AP$(x_g,y_g,w_g,h_g)$8(16–32 px), and AP$(x_g,y_g,w_g,h_g)$9(32–64 px). VisDrone-2019 contains 261 908 video frames and 10 209 stills across 14 cities, and COCO-2017 contains approximately 330 000 images with 80 classes (Guan et al., 3 Jan 2026).

On AI-TOD-v2, the reported progression is base FCOS at $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$0 AP, DW at $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$1 AP, RFAssigner at $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$2 AP, and RFAssigner* at $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$3 AP, which is described as a new SOTA. The gains are most pronounced on very tiny and tiny scales, with AP$$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$4 changing from $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$5 to $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$6 and AP$$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$7 from $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$8 to $$N_g:\; \mu_g= \begin{pmatrix} x_g\ y_g \end{pmatrix}, \quad \Sigma_g= \begin{pmatrix} (w_g/2)^2 & 0\ 0 & (h_g/2)^2 \end{pmatrix}.$$9. On VisDrone, DW improves from $N_e(\mu_e,\Sigma_e)$0 AP to $N_e(\mu_e,\Sigma_e)$1 AP under RFAssigner. On COCO, DW improves from $N_e(\mu_e,\Sigma_e)$2 AP to $N_e(\mu_e,\Sigma_e)$3 AP, while AP$N_e(\mu_e,\Sigma_e)$4 increases from $N_e(\mu_e,\Sigma_e)$5 to $N_e(\mu_e,\Sigma_e)$6. Compared with DW+RFLA, ATSS, and AutoAssign, RFAssigner is reported as the only method that consistently raises small-object AP without sacrificing medium or large objects (Guan et al., 3 Jan 2026).

The ablation studies further characterize the receptive-field criterion. On AI-TOD-v2, using WD gives $N_e(\mu_e,\Sigma_e)$7 AP, KLD gives $N_e(\mu_e,\Sigma_e)$8 AP, NWD gives $N_e(\mu_e,\Sigma_e)$9 AP, and GCD gives $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$00 AP. For ambiguous matching, the best thresholds are $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$01, which produce $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$02 AP; setting $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$03 too low admits noise and decreases AP by $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$04, while setting $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$05 too high or disabling ambiguous matching by $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$06 loses hard positives and also decreases AP by $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$07. Across benchmarks, RFAssigner improves AP$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$08 by $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$09–$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$10 and AP$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$11 by $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$12–$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$13 while AP$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$14 remain unchanged or slightly improved (Guan et al., 3 Jan 2026).

6. Limitations, impact, and future directions

The original RFLA identifies several limitations of its distance choices and receptive-field model. WD is not scale-invariant, which is why the authors prefer KLD under large scale variance, while KLD can be ill-defined when distributions barely overlap. Jensen–Shannon divergence was discarded because it had no closed form. The method also relies on a fixed Gaussian ERF model, although real networks may have non-Gaussian tails or anisotropies. Proposed extensions include learning per-layer covariance shape, learning the distance metric itself, combining the two-stage ranking with optimal-transport or soft-matching in very high-density scenes, and extending RFD from a purely geometric prior to a joint feature-similarity prior using predicted scores or feature-map activations (Xu et al., 2022).

RFAssigner states a different set of limitations. It requires minor extra memory during training to store RFD scores and is currently tied to FCN architectures because of TRF computation. The future directions listed for that formulation are extension to transformer backbones, dynamic learning of RFD thresholds, and per-class or adaptive $$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$15 in the top-$$G_n(x,y)=\frac{1}{2\pi\sigma_x\sigma_y} \exp\Bigl(-\tfrac{(x-x_n)^2}{2\sigma_x^2} -\tfrac{(y-y_n)^2}{2\sigma_y^2}\Bigr),$$16 rule. The same work presents RFAssigner as the first generic, soft-assignment strategy that combines point priors with adaptive GRF prior supplementation, fully differentiable and inference-free, and attributes its empirical profile to alleviating the positive-sample starvation of prior soft assignment while avoiding the over-specialization of hard GRF-only schemes (Guan et al., 3 Jan 2026).

A recurring misconception is that receptive-field label assignment is merely a change of distance metric. The two papers indicate a broader change in training geometry: the positive set is no longer determined solely by overlap thresholds or center rules, but by how a feature location’s modeled receptive field corresponds to a ground-truth object. In the original RFLA this leads to hierarchical hard assignment optimized for tiny-object recall; in RFAssigner it becomes adaptive supplementation within dual-weight soft assignment. This suggests that the main significance of RFLA lies not only in Gaussian modeling, but in redefining what constitutes a valid positive sample across object scales.