---
title: Dynamic Anchor Selection (DAS)
url: https://www.emergentmind.com/topics/dynamic-anchor-selection-das
type: topic
---

# Dynamic Anchor Selection (DAS)

Dynamic Anchor Selection (DAS) denotes a family of adaptive mechanisms that replace fixed anchors with anchors chosen, generated, or weighted as a function of local geometry, data distribution, temporal state, or task context. Across recent work, the anchor object varies widely: it can be a subset of UWB anchor pairs in TDOA localization, a set of positive training anchors in object detection, a query vector in video segmentation, an anchor line in rail detection, a prompt token in CLIP adaptation, a structural delimiter in diffusion LLM infilling, a cluster exemplar in generalized category discovery, or a subset of Gaussians in streaming reconstruction. The unifying principle is that anchor choice is treated as an optimization variable rather than a static heuristic, typically with the goal of reducing localization error, improving assignment quality, stabilizing training, or satisfying compute and format constraints [2404.05067] [1912.02424] [2404.00086] [2603.17227].

## 1. Scope of the concept across research domains

In the cited literature, DAS does not refer to one canonical algorithm. It is instead a recurring design pattern in which a downstream estimator or learner is conditioned on anchors that are selected adaptively rather than prescribed a priori. The adaptive signal may come from calibration error, confidence statistics, density, clustering structure, contextual activation, reinforcement-learning reward, or a model’s own generation confidence [2404.15330] [2012.04150] [2510.16342] [2512.12925].

| Domain | Anchor object | Selection target |
|---|---|---|
| UWB TDOA localization | Anchor pairs or anchor subsets | RMSE, doorway transition cost, LOS/NLOS-aware feasibility |
| Object detection and tracking | Training anchors, sparse anchors, query anchors | Positive/negative assignment, localization potential, emergence/disappearance handling |
| Structured generation and prompt learning | Prompt anchors, end anchors, anchor tokens | Stability–plasticity balance, format compliance, task adaptation |
| Open-world learning and erasure | Cluster exemplars, semantic anchors | Hard pseudo-label purity, erasure robustness, boundary preservation |
| 4D Gaussian streaming | Anchor budget and Gaussian subset | PSNR–runtime trade-off |

This breadth matters because it prevents an overly narrow reading of the term. In some papers, “dynamic” means per-image or per-object training-time assignment; in others it means per-zone runtime switching, per-epoch exemplar refresh, or sequential policy selection under a budget. A plausible implication is that DAS is best understood as an optimization schema over anchor choice rather than as a specific architectural primitive.

## 2. Core mathematical patterns

Although the domains differ, the mathematical structure of DAS is usually one of a few recurring forms. One common form is empirical error minimization over a candidate pool. In the UWB RMSE-based formulation, for each zone $z$ and candidate anchor-pair set $S$, the calibration objective is

$$
\mathrm{RMSE}_{z}(S) = \sqrt{\frac{1}{K_z}\sum_{k\in \mathrm{zone}\ z} \|\hat p_{S,k} - p_{\mathrm{ref},k}\|^2},
$$

followed by the selection rule $S^\star_z=\arg\min_S \mathrm{RMSE}_z(S)$ [2404.05067]. The doorway-based UWB variant uses a different calibration objective, the “door transition” cost,

$$
\mathrm{Cost}(\mathrm{TDOA\;set})=\sum_{i=1}^n (h_i+v_i),
$$

with separate minimizers for entering and exiting a zone [2404.15330].

A second form is adaptive thresholding from local statistics. ATSS computes per-ground-truth IoU statistics on nearby anchors and defines a dynamic threshold

$$
t_i^k = \mu_i^k + \sigma_i^k,
$$

then marks as positive the candidate anchors whose IoU exceeds $t_i^k$ and whose centers fall inside the ground-truth box [1912.02424]. The anchor criterion is therefore not fixed globally, but estimated from the sample distribution induced by each object and feature-pyramid level.

A third form is regression-aware anchor quality scoring. Dynamic Anchor Learning defines the matching degree

$$
\mathrm{md}(A)=\alpha\cdot sa + (1-\alpha)\cdot fa - u^\gamma,
$$

where $sa$ is pre-regression IoU, $fa$ is post-regression IoU, and $u=|sa-fa|$ is a regression-uncertainty penalty [2012.04150]. Here DAS is neither pure geometry nor pure confidence; it is a composite quality functional designed to reduce the divergence between classification confidence and localization accuracy.

A fourth form is policy optimization over subsets and budgets. Efficient Gaussian Streaming models anchor choice as an action $a_t=(\kappa_t,\Omega_t)$, where $\kappa_t\in\{256,512,\dots,8192\}$ is the anchor budget and $\Omega_t$ is the selected subset, and optimizes expected reward with REINFORCE under a quality–runtime trade-off [2603.17227].

These formulations show that “anchor” is a role, not a datatype. The anchor may be a measurement pair, a box prior, a query, a prompt token, or a sampled primitive; what makes the method a DAS method is the adaptive selection mechanism tied to a downstream loss or utility.

## 3. UWB TDOA localization and zone-conditioned selection

The most literal use of DAS in the supplied literature appears in UWB TDOA indoor localization. In "Adaptive Anchor Pairs Selection in a TDOA-based System Through Robot Localization Error Minimization" [2404.05067], the method is explicitly split into an off-line calibration phase and an on-line positioning phase. The coverage area is partitioned into non-overlapping zones, typically along natural room boundaries or corridors. During calibration, a six-wheeled robotic chassis (Dagu Wild Thumper 6WD) carrying a UWB tag on a 1.4 m pole and a Scanse Sweep LiDAR is driven through all zones. LiDAR scans are matched via GraphSLAM + ICP to produce a smooth reference trajectory with subdecimeter accuracy $(\sigma\approx 3\ \mathrm{cm})$. For each timestamp and each candidate anchor-pair set, a Levenberg–Marquardt nonlinear least-squares solver estimates position from TDOA residuals, and the per-zone RMSE is computed against the LiDAR reference. With $M=6$ anchors, there are $C(M,2)=15$ possible pairs, and the reported real experiment tested all combinations of $3,4\ldots$ up to $5$ pairs, with no anchor appearing in more than three pairs per set, yielding $4\,487$ candidate sets in the real experiment and $547$ in simulation. Online, the latest position estimate identifies the current zone, the system switches to the corresponding preselected pair set, and an Unscented Kalman Filter fuses the LS pose, velocity models, and prior state. In a 5-zone simulated apartment with $M=6$ anchors and $\sigma_{\mathrm{TOA}}=0.6\ \mathrm{ns}$, using a fixed set of 4 TDOA pairs across all zones yielded an 80th-percentile error of approximately $45\ \mathrm{cm}$ and maximum approximately $1.5\ \mathrm{m}$, whereas DAS reduced these to approximately $30\ \mathrm{cm}$ and approximately $0.8\ \mathrm{m}$. In the dynamic person-tracking experiment, DAS yielded a median trajectory error of approximately $25\ \mathrm{cm}$ and a 90th-percentile of approximately $50\ \mathrm{cm}$ [2404.05067].

"Anchor Pair Selection in TDOA Positioning Systems by Door Transition Error Minimization" [2404.15330] reframes the selection problem around doorway geometry rather than global zone RMSE. The floor-plan is overlaid with an occupancy-grid map with resolution around $5\,\mathrm{cm/pixel}$, and detected doors partition the free space into disjoint zones. A Single Shot Detector with RetinaNet backbone (ResNet-50 v1 FPN) is run on a $640\times 640$ occupancy-grid image to regress doorway bounding boxes, from which the midpoint $p_d$, orientation $\theta_d$, and half-width $w_d$ are recorded. The calibration prior is that a user crossing a doorway must lie on the door axis, and the selected anchor-pair set minimizes the accumulated off-axis and along-axis projection error at door crossings. The method determines anchor sets separately for entering and exiting each zone to account for body shadowing. In a $48\,\mathrm{m}^2$ furnished apartment with 6 UWB anchors and 1 body-worn tag, using DAS produced median trajectory error of approximately $0.32\,\mathrm{m}$, compared with approximately $0.72\,\mathrm{m}$ for a single static anchor-set baseline, corresponding to an improvement of approximately $0.4\,\mathrm{m}$ on average and approximately $35$–$40\%$ reduction in error [2404.15330].

"Dynamic Anchor Selection and Real-Time Pose Prediction for Ultra-wideband Tagless Gate" [2402.17778] uses yet another criterion: LOS/NLOS classification from UWB channel impulse response and IMU data. Each anchor produces an eCIR vector of approximately 135 samples instead of the full 1,016-sample CIR, reducing USB latency from $223\ \mathrm{ms}$ to $17.8\ \mathrm{ms}$. A four-block Conv1D network with Instance Normalization, ReLU, Dropout(0.2), MaxPool1D, and an exponentially weighted moving average with $\alpha=0.8$ performs binary LOS/NLOS classification, achieving end-to-end accuracy $0.984$. DAS then greedily searches anchor subsets of size 4, ordered by ascending predicted NLOS cost, subject to a convex-hull constraint that the previous mobile-device estimate lies inside the selected anchors. In a $9\times 9\ \mathrm{m}^2$ office with six anchors, the NLOS mean error fell from $10.01\ \mathrm{cm}$ to $3.78\ \mathrm{cm}$, reported as $62\%$ improvement, while LOS mean error improved from $2.64\ \mathrm{cm}$ to $2.23\ \mathrm{cm}$; the same system reports four-pose classification accuracy $0.961$ on a Samsung Galaxy Note20 Ultra and Qorvo UWB board [2402.17778].

Taken together, these localization papers illustrate three distinct DAS logics: empirical per-zone calibration, geometry-constrained doorway calibration, and per-epoch channel-aware subset selection. The common outcome is that fixed anchor-pair sets are replaced by situationally preferable subsets.

## 4. Detection, segmentation, and geometry-aware vision anchors

In object detection, DAS often appears as adaptive sample assignment rather than explicit runtime subset switching. "Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection" [1912.02424] argues that the essential difference between anchor-based and anchor-free detectors is how positive and negative training samples are defined. ATSS selects the $N$ anchors closest to each ground-truth center on each FPN level, computes the mean and standard deviation of their IoUs, and uses $\mu+\sigma$ as a dynamic threshold. On MS COCO, the reported RetinaNet baseline with GroupNorm, GIoU loss, and centerness improves from $37.0$ AP to $39.3$ AP with ATSS; the paper further reports test-dev results up to $50.7\%$ AP with multi-scale test and no introduced overhead [1912.02424]. This formulation is historically important because it casts anchor adaptation as a label-assignment problem rather than a box-design problem.

For arbitrary-oriented detection, "Dynamic Anchor Learning for Arbitrary-Oriented Object Detection" [2012.04150] replaces IoU-only assignment with the matching degree that combines input IoU, output IoU, and a regression-uncertainty penalty. Positive anchors satisfy $\mathrm{md}(A)>\tau_{\text{pos}}$, with $\tau_{\text{pos}}=0.6$ in the paper, and the same matching degree also produces compensation weights in classification and regression losses. Reported gains include HRSC2016 AP$_{50}$ from $80.8\%$ to $88.6\%$, DOTA mAP from approximately $68.4\%$ to approximately $71.4\%$ on the R50 baseline, and Pascal VOC 2007 mAP from $74.9\%$ to $76.1\%$; the paper emphasizes that this occurs without extra inference cost and with only a few horizontal preset anchors [2012.04150].

In video segmentation, "DVIS-DAQ: Improving Video Segmentation via Dynamic Anchor Queries" [2404.00086] generalizes the anchor notion from boxes to query states. A single static background query is replaced by dynamic anchor queries generated per candidate object, with separate mechanisms for emergence and disappearance. Emergence anchors use current-frame proposal features, while disappearance anchors use momentum-weighted historical appearance features and the closest initial segmenter query in cosine similarity. The method also introduces query-level object Emergence and Disappearance Simulation so that rare appearance events are synthesized without extra forward cost. The reported overhead is 2%, while OVIS val AP increases from $30.2$ for the DVIS baseline to $37.1$ with DAQ + ES + DS, and YouTube-VIS ’19 AP rises from $51.2$ to $55.2$ [2404.00086].

Rail and lane detection provide a geometric reinterpretation. "DALNet: A Rail Detection Network Based on Dynamic Anchor Line" [2308.11381] uses an anchor line parameterized by $(x_{\rm start},y_{\rm start},\theta)$ rather than a bounding box. A dynamic anchor line generator predicts a heatmap, offset map, and slope map from FPN level $P_4$ to generate image-adaptive anchor lines for each rail instance. On the DL-Rail dataset, which contains 7000 pairs of images and annotations, the full system with PPM global context reports F1@50 $96.43$, F1@75 $65.48$, mF1 $59.79$, and $212$ FPS; on TuSimple it reports $96.68\%$ accuracy, and on LLAMAS it reports F1 $96.12\%$ [2308.11381].

Sparse transformer detection offers yet another anchor interpretation. "ASAG: Building Strong One-Decoder-Layer Sparse Detectors via Adaptive Sparse Anchor Generation" [2308.09242] predicts dynamic anchors on patches rather than grids, adaptively selecting which feature maps and which locations to probe. It couples this with Query Weighting to mitigate training instability under adaptive sampling. On COCO val with a ResNet-50 backbone, ASAG-S with approximately 100 anchors reports AP $43.9$ at $30.1$ FPS, and with approximately 300 anchors reports AP $45.0$ at $29.2$ FPS, outperforming dense-initialized one-layer variants in the reported speed–accuracy trade-off [2308.09242].

These vision papers collectively show that DAS can modify at least four distinct loci in a perception stack: target assignment, query initialization, geometric priors, and sparse proposal generation. The anchor is dynamic not because it moves physically, but because the system estimates which reference structure is appropriate for the current sample.

## 5. Prompt anchors, structural anchors, and continual adaptation

In prompt-based and generative systems, DAS is applied to symbolic or embedding-level anchors. "Dynamically Anchored Prompting for Task-Imbalanced Continual Learning" [2404.14721] maintains a single general prompt $p_g$ across the task stream and regularizes it with two task-dependent prompt anchors: a boosting anchor $p_t^b$, learned by fitting a task-specific prompt on the current task, and a stabilizing anchor $p_t^s$, defined as an online centroid of previous boosting anchors weighted by task size. The alignment term is cosine-based, $\mathcal{L}_a(p_g,p)=1-\frac{p_g^\top p}{\|p_g\|\|p\|}$, and the interpolation coefficient $\lambda_t$ depends on task size so that smaller tasks emphasize plasticity and larger tasks emphasize stability. The paper reports absolute improvements of roughly $+4.5\%$ to $+15\%$ over prior rehearsal-free prompt methods under task-imbalanced settings; on shuffled TICL-CIFAR100, DualPrompt achieves $49.5\%$ average accuracy, CODA-Prompt $60.9\%$, and DAP $61.4\%$ [2404.14721].

"Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models" [2606.04535] treats begin- and end-anchor tokens as structural delimiters for reasoning or JSON generation. The method is training-free: for each masked block, it prepends a begin anchor, runs one diffusion infer step, scans the predictions for an end anchor or its prefix above confidence threshold $c$, and either truncates or extends the masked block by $\Delta$ until an end anchor is found or the hard limit $M$ is reached. Stage I therefore localizes the proper span endpoint, and Stage II performs iterative infilling inside the chosen boundaries. On zero-shot GSM8K, the reported format score increases from $58.83\%$ to $72.63\%$ and accuracy from $14.86\%$ to $46.78\%$; on MATH, format score rises from $29.10\%$ to $76.82\%$; on WikiBio JSON generation, valid JSON reaches $79.84\%$ with hallucination score $0.15\%$ [2606.04535]. The paper explicitly presents DIA as an instance of DAS in diffusion LLMs.

"AnchorOPT: Towards Optimizing Dynamic Anchors for Adaptive Prompt Learning" [2511.21188] extends prompt anchors further by making both anchor values and anchor positions learnable. Instead of handcrafted explicit textual tokens such as “shape” or “color,” it introduces continuous anchor embeddings and a learnable position matrix optimized with Gumbel-Softmax. Training is two-stage: Stage I learns anchor embeddings by MSE alignment to LLM-generated class descriptions; Stage II freezes the anchors and optimizes soft tokens plus the position matrix under a cross-entropy plus KL objective. Reported gains on 11 base-to-novel tasks include raising the harmonic mean of CoOp from approximately $71.7$ to approximately $78.7$, with similar gains for CoCoOp, MaPLe, and DePT [2511.21188].

These works broaden the anchor concept from geometric reference to semantic constraint. The anchor may delimit where content should end, indicate how prompts should remain stable across skewed task streams, or provide a data-learned prompt token whose value and position are themselves adaptive variables.

## 6. Cluster exemplars, semantic erasure anchors, and budget-aware streaming

In open-world recognition, DAS becomes exemplar mining. "Sharpness-aware Dynamic Anchor Selection for Generalized Category Discovery" [2512.12925] selects anchor samples for unknown classes at the end of each epoch, using a combination of cluster density and classifier confidence. For each novel cluster, it computes a truncation threshold
$T_{\mathrm{tru}}=p_{\mathrm{avg}}+\omega(p_{\mathrm{max}}-p_{\mathrm{avg}})$ over maximum class probabilities, sets a uniform anchor count by a $\gamma$-quantile over per-cluster counts, identifies the density peak by inverse average $k$-NN distance, and then picks the top-$\eta$ candidates near that peak by confidence. These anchors are injected into the labeled pool with one-hot labels for the next epoch. With SimGCD as baseline, DAS alone raises ImageNet-100 new-class ACC from $77.9$ to $80.7$ and CUB new-class ACC from $57.7$ to $66.2$; combined with Loss Sharpness Penalty, new-class ACC reaches $83.7$ on ImageNet-100 and $68.9$ on CUB [2512.12925].

In concept erasure for diffusion models, DAS becomes semantic anchor mining. "Beyond Fixed Anchors: Precisely Erasing Concepts with Sibling Exclusive Counterparts" [2510.16342] defines Sibling-Exclusive Concepts as anchors that share the same parent in a semantic hierarchy as the target concept while having disjoint core attributes. SELECT evaluates candidates in two stages: Stage I uses the Contextual Activation Score, via the Context-Raise-Ratio $U_c$, to retain anchors with low residual activation under contextual clues; Stage II uses the Semantic Coherence Score, defined as a perplexity ratio $\mathrm{CoS}(c_i)=\frac{\mathrm{PPL}(p)}{\mathrm{PPL}(p')}$, to pick the most contextually natural anchor. The paper reports average anchor mining time of only 4 seconds for one concept, object-erasure harmonic mean $H_o\approx 87.3$ for SELECT versus approximately $56.9$ for a fixed “general” anchor, and in NSFW erasure a drop in average NudeNet detection sum from approximately $75$ to $15$ while maintaining CLIP scores and FID [2510.16342].

In dynamic scene reconstruction, DAS becomes policy-based subset selection under a compute budget. "Adaptive Anchor Policies for Efficient 4D Gaussian Streaming" [2603.17227] replaces Farthest Point Sampling at fixed $K=8192$ anchors with a learned policy that jointly chooses the anchor budget and the anchor subset. Candidate Gaussians are encoded by normalized coordinates, opacity, log average scale, and distance from the centroid, then processed by a pointwise MLP and a 2-layer Set Transformer. A budget head predicts over $\{256,512,1024,\dots,8192\}$ and a selection head samples a subset without replacement. On unseen data in fast rendering mode at 256 anchors, EGS improves PSNR by $+0.52$--$+0.61\ \mathrm{dB}$ while running $1.29$--$1.35\times$ faster than IGS@8192 on N3DV and MeetingRoom. The same-budget ablation reports EGS–FPS gains of $+0.424\ \mathrm{dB}$ at 256 anchors, $+0.479\ \mathrm{dB}$ at 512, and $+0.458\ \mathrm{dB}$ at 1024 [2603.17227].

This cluster of papers shows that DAS can be used to improve supervision quality, preserve semantic boundaries, or allocate rendering resources. The anchor is not merely a support point; it can also be a high-purity pseudo-label carrier, a semantically exclusive counterpart, or a budgeted rendering primitive.

## 7. Recurring design choices, misconceptions, and limitations

A frequent misconception is that DAS is synonymous with anchor-free design. The ATSS literature explicitly argues otherwise: if anchor-based and anchor-free detectors adopt the same definition of positive and negative samples during training, there is no obvious difference in final performance, and the key factor is how training samples are selected [1912.02424]. DAS therefore often modifies an anchor-based system rather than replacing anchors altogether.

Another misconception is that “dynamic” always implies runtime switching. The supplied papers show at least four update cadences: off-line calibration with lookup-table deployment in UWB localization, per-image assignment in object detection, per-epoch anchor refresh in generalized category discovery, and one-step policy selection in streaming reconstruction [2404.05067] [2512.12925] [2603.17227]. A plausible implication is that the most informative taxonomy of DAS is by update schedule and objective, not by application area alone.

The limitations reported in the literature are also heterogeneous but patterned. Door-transition DAS requires a reasonably accurate occupancy grid and door detection, and open-plan or cluttered layouts may yield spurious zones [2404.15330]. DynaPose depends on reliable LOS/NLOS classification and uses buffer-based outlier handling to stabilize anchor selection [2402.17778]. ASAG introduces several hyperparameters for probing thresholds and patch sizes [2308.09242]. DIA relies on manual anchors with fixed semantic roles and adds iterative expansion overhead that may not suit ultra-low-latency scenarios [2606.04535]. EGS adds a small latency of less than 40 ms due to the Transformer sampler and is trained primarily in fast rendering mode with a frozen backbone [2603.17227]. SELECT depends on candidate generation and semantic evaluation pipelines, even though the reported wall-clock anchor-mining time is 4 seconds per concept [2510.16342].

A broader pattern nevertheless emerges. DAS methods usually combine three ingredients: a candidate pool, an adaptive quality signal, and a fallback or stabilization mechanism. In UWB systems the fallback is zone lookup or convex-hull feasibility; in DAL it is the highest matching-degree anchor when no anchor exceeds threshold; in GCD it is uniform anchor count across novel clusters; in continual learning it is the stabilizing centroid prompt; in ASAG it is Query Weighting. This suggests that successful DAS rarely depends on dynamic scoring alone; it also depends on a mechanism that prevents degenerate anchor choices when the adaptive signal is noisy.

Across the supplied literature, the significance of DAS lies less in any one metric than in the repeated demonstration that fixed anchors are often mismatched to local conditions. Whether the problem is NLOS body shadowing, per-object IoU variation, rare emergence events, task imbalance, semantic leakage in concept erasure, or over-allocation of Gaussian primitives, dynamic anchor choice is used to make the reference structure itself contingent on the instance being processed.

Source: https://www.emergentmind.com/topics/dynamic-anchor-selection-das