Abstract
Test-time adaptation (TTA) can improve detector robustness under distribution shift, but indiscriminate online updating may amplify errors from unreliable frames. We present AdaTent++, a reliability-guided hybrid TTA framework for traffic sign detection. For each unlabeled frame, the framework combines predictive entropy and cross-view agreement to select among a smoothed batched normalization (BN)-statistics update, a guarded BN-affine gradient update, and no adaptation. Exponential moving average (EMA) threshold tracking, prototype regularization, gradient control, and state restoration are used as supporting safeguards rather than claimed as independent novelties. On one public traffic sign dataset evaluated under 12 corrupted domains, AdaTent++ achieves a mean average precision (mAP50) of 68.87%, compared with 68.03% for the strongest baseline, T3A, corresponding to an improvement of 0.84 percentage points. It ranks first in three of the twelve domains and is less effective under several structure-destructive corruptions. The results indicate a favorable average trade-off in the studied setting, while cross-dataset, cross-architecture, embedded-device, and long-horizon validation remain open.
1 Introduction
Traffic sign recognition (TSR) is a key perception module in autonomous driving and advanced driver assistance systems (ADAS), providing real-time enforcement of rules such as speed limits, yielding, and prohibitions. However, in real-world road environments (), signs commonly present visual challenges such as small distant targets and glare/occlusion. Pronounced cross-regional differences in shape, color, and semantics coupled with weather, illumination, and temporary roadworks continually introduce distribution shifts. Although current detection models perform stably within the training distribution (source domain), their recognition performance and confidence calibration tend to degrade significantly under cross-domain settings, sensor aging, or compression noise (). Relying solely on offline retraining suffers from data collection/annotation latency, struggles to cover short-term changes in time, and incurs high maintenance costs. To alleviate these issues, researchers have proposed a variety of traditional test-time adaptation (TTA) methods, such as adaptive batch normalization (AdaBN) based on statistics alignment and Tent based on entropy minimization, which enable lightweight updates during inference without labels and thereby improve cross-domain robustness to some extent. Nevertheless, in TSR scenarios, these methods lack effective modeling of uncertainty and multi-view consistency and are sensitive to small distant objects and strong noise. Their stability is difficult to guarantee under edge-compute and latency constraints ().
Conventional TTA for TSR (e.g., AdaBN for statistical alignment and Tent for entropy minimization) exhibits notable limitations (). First, it lacks reliable modeling of sample uncertainty and multi-view consistency, making it difficult to distinguish high- and low-confidence and noisy samples and thus prone to harmful updates (). Second, the objectives are biased toward classification with insufficient constraints on bounding-box regression, leading to semantic–geometric inconsistency and localization drift. Third, under small-batch/streaming inference and distribution shifts, the method is susceptible to BN drift and parameter oscillation. Fourth, it remains vulnerable to small distant objects and strong noise and lacks stability safeguards such as adaptive thresholding, gradient regularization, and state rollback, making long-term stable operation within edge latency budgets difficult.
To this end, we propose AdaTent++, which is an uncertainty–consistency-oriented routing scheme that enables sample-level differentiated updates (high-confidence samples take a lightweight AdaBN path, uncertain samples perform “entropy minimization + prototype alignment,” and noisy samples are safely skipped), together with exponential moving average (EMA)-based thresholding, gradient regularization, dynamic step scheduling, and batch normalization (BN)-state protection, thereby achieving more robust and real-time test-time adaptation under label-free, compute-constrained settings ().
To validate the above design and ensure its effectiveness under in-vehicle constraints, we conducted a systematic empirical study of AdaTent++ in a label-free online setting. Specifically, we chose the lightweight and efficient EfficientDet-D0 as the base model to meet edge-compute and latency constraints and to minimize architectural confounds. We evaluated across a variety of distribution-shift scenarios covering both synthetic and real-world perturbations. Compared with state-of-the-art TTA baselines, AdaTent++ achieves significant improvements on aggregate metrics (e.g., mean average precision, or mAP) and remains superior in most specific corruption/domain-shift settings, indicating stronger robustness and adaptability for traffic sign detection.
Our main contributions can be summarized as follows:
We formulate a reliability-guided test-time adaptation policy for each frame that selects among no update, a statistics-only update, and a guarded gradient update.
We provide a resource-conscious realization of this policy using EMA threshold tracking, conservative BN-affine optimization, and state protection as supporting safeguards.
We evaluate the resulting framework on one public traffic sign dataset under 12 corrupted domains and report both its average improvement and its domain-specific limitations.
In real-world road scenes, distribution shifts continually erode detector performance, motivating the community to explore label-free test-time adaptation (TTA) paradigms that correct model bias online during inference. Along this line, proposed AdaBN, which alleviates the immediate mismatch from source-domain bias by aligning normalization statistics at inference. However, it is insensitive to sample reliability and is prone to statistical drift in noisy settings and under small-batch/streaming inference. Subsequently, introduced Tent, which drives online updates via entropy minimization and further improves cross-domain adaptability. Yet under heavy corruptions, occlusions, and small objects, it tends to amplify erroneous confidence, and the lack of unsupervised constraints on the localization head leads to semantic–geometric inconsistency. To enhance stability, proposed CoTTA. proposed EATA, and proposed SAR, employing teacher–student schemes, explicit regularization, and memory mechanisms to suppress catastrophic drift (). However, these methods still lack controllable sample-level update decisions and incur nontrivial overhead under latency/compute budgets, limiting deployment. Detection-oriented refinements (e.g., updating only a small subset of BN/head parameters or incorporating multi-view consistency and prototype alignment) partially compensate for missing supervision but remain insufficient to answer “when to update, which samples to update, and how to remain stable over time” ().
We propose AdaTent++ to address the above limitations. The method introduces, at the sample level, an uncertainty- and consistency-oriented dynamic routing mechanism to determine whether and how to update and incorporates EMA-based adaptive thresholds to stabilize decision boundaries. For uncertain samples, we design a unified hybrid branch that jointly applies entropy minimization and prototype alignment to mitigate the mismatch between classification and localization and to enhance noise robustness. Meanwhile, we adopt a BN “update–freeze–pullback or rollback” procedure with a safe-skip strategy as stability guardrails, thereby achieving robust and controllable test-time adaptation under label-free, edge-compute, and latency constraints.
2 Materials and methods
2.1 Method overview
Given an unlabeled test frame, AdaTent++ first computes a low-cost reliability estimate from predictive entropy and cross-view agreement. The reliability estimate then selects one of three mutually exclusive actions: a statistics-only BN update, a guarded BN-affine gradient update, or no adaptation. The following subsections define the reliability measure, routing rule, adaptation actions, and state-protection procedure without repeating the application motivation. As shown in Figure 1, we unlock only the affine parameters of BN before inference and back up all BN states. For each unlabeled frame, we first generate several views and, at the highest-confidence anchor, compute the uncertainty , the multi-view consistency , and the noise score . EMA-smoothed thresholds with hysteresis implement one-shot routing. When exceeds the threshold, adaptation is directly skipped. For high-confidence samples, we adopt a lightweight AdaBN path and blend BN statistics with a safety-margin coefficient. The remaining samples enter Tent++, which, with BN statistics frozen, performs 1–2 dynamic steps of temperature-scaled entropy minimization together with prototype alignment under gradient control. All paths output detection results, while routing ratios and thresholds are logged, and BN states are periodically restored to constrain the adaptation magnitude and ensure long-term stability. The proposed framework integrates two primary adaptation routes: a lightweight AdaBN path for high-confidence samples, and a highly augmented entropy-minimization branch for uncertain samples. Because this second branch significantly enhances the standard Tent baseline by incorporating prototype alignment and gradient regularization (which we denote as the Tent++ branch), the overall unified hybrid framework is named AdaTent++.
FIGURE 1
2.2 BN-only and steady-state anchor
In noisy-frame and small-batch streaming settings, whole-model online updates are easily destabilized by gradient noise. AdaBN often overwrites BN running statistics, allowing a single frame to tug the statistics and cause drift (). Tent immediately minimizes entropy on every frame, even when the sample is unreliable, which pushes parameters toward overconfidence ().
To reduce instability and latency simultaneously, we shrink the updatable region to BN-affine parameters . We attach a lightweight optimizer only to them (Adam, learning rate (LR) ), fully freeze all other layers, and back up the running mean/variance of all BN layers as a steady-state reference before inference.
This design establishes three guard rails. (i) The update surface is tiny, which limits gradient-noise injection and fits edge budgets. (ii) BN-affine terms provide a first-order adaptation channel that is sufficient to compensate most domain shifts without full backpropagation through the network. (iii) Any suspect write can be undone via pullback or rollback to the backed-up steady state.
Unlike AdaBN’s “overwrite first,” statistics are not rewritten blindly. Unlike Tent’s “update first,” updates are deferred until reliability is tested, which avoids writing incorrect confidence into the model.
Two invariants ensure the rigor of the procedure. First, except for the affine parameters of BN, no parameters are trainable. Second, whenever statistics must be frozen, BN layers should remain in eval mode, with affine transformations applied and no updates to the running statistics. Violating either invariant introduces hidden coupling and causes statistics to drift during backpropagation. The backup and restore overhead scales linearly with the number of BN layers, requires only a small constant amount of memory and I/O, and is far below the cost of a single backpropagation step, thereby making batch- or session-level windowed control practical.
2.3 Multi-view reliability and dynamic routing (EMA + hysteresis, contrasts, and operational definitions)
Prior to any adaptation, the question, “does this frame warrant an update?” must be answered. This step is missing in AdaBN (statistics-only) and Tent (entropy only). For each frame , a set of views is generated (default ) using lightweight augmentations drawn from general, low-light, blur, and weather, and a forward pass is performed on each view (). Subsequently, at the highest-confidence anchor, the class distribution and the prediction are read to avoid background-anchor noise. Finally, uncertainty is quantified with entropy, consistency is measured with majority voting, and the two are coupled into a noise score (Equations 1–3).
Intuitively, when “high + low ” occurs, the observation lies near a class boundary, and the views disagree. Therefore, the observation is unreliable and should not be written. Conversely, when “low + high ” occurs, the semantics are clear and cross-view agreement holds, and it is, therefore, reliable and can serve as evidence for lightweight adaptation. Compared with AdaBN’s “always align statistics” and Tent’s “always reduce entropy,” uncertainty consistency becomes the primary gating criterion.
Fixed thresholds jitter under changing scenes, so gates must track the environment. EMA thresholds are maintained with per-frame rate limiting and value clipping (Equation 4):
Hysteresis is added to avoid boundary ping-pong (
Equation 5),
and locked with the previous path. One routing decision is made per frame:
if (empirically ), skip;
if and (or persistently satisfies ), choose high-confidence lightweight;
otherwise, choose robust-under-uncertainty.
Under this router, AdaBN would rewrite statistics where we may do nothing, and Tent would push entropy where we may only blend BN lightly, tying “whether and how to move” strictly to reliability. Operational notes to ensure rigor are as follows. First, the upper bound of is constrained by the latency budget. Second, the “highest-confidence anchor” refers to the anchor with the largest across-class posterior on that view, rather than the anchor with the most detected boxes. Finally, when a tie occurs in majority voting, the tie is broken using the sum of posteriors across views to avoid accidental flips caused by discrete voting.
2.4 Three-path mechanics and stability domain (expanded details, risk control loop, and interleaved contrasts)
Path A (high confidence, zero backprop). Overwrite-style alignment directly writes small-batch fluctuations into running stats, causing jitter/drift. Smoothed blending is used with strength set by a safety margin:
From one forward pass on the current frame, “new statistics” are sampled and then blended with the backed-up “old” values:
Safer frames larger to absorb more of the live distribution. Borderline frames remain conservative. Compared with AdaBN’s hard overwrite, this path achieves zero backpropagation and near-zero latency, thereby markedly reducing single-frame pull (Equations 6, 7). Relative to Tent’s per-frame gradient updates, it avoids both computational overhead and gradient uncertainty. BN train/eval toggling is crucial: to obtain “new statistics,” BN must temporarily allow updates to the running statistics, and it should switch back to eval immediately after blending to prevent further disturbance.
Path B (robust under uncertainty, minimal backprop). Only in the region not skipped yet not safe enough do we run minimal backprop. Unlike Tent’s entropy-only objective, the loop is closed with steady start
small steps
hybrid objective
risk gate:
Steady start. First, perform a single pre-blend of the BN statistics with a fixed to place the starting point closer to the current distribution. Subsequently, set BN to eval to freeze the running statistics during backpropagation, in order to avoid the hidden coupling that arises from updating statistics while computing gradients.
Small steps. Stack the views into a micro-batch and perform dynamic 1–2 step updates. When is high or is low (far from the thresholds), execute one step to cap latency and risk. Otherwise, execute two steps.
Hybrid objective. Employ temperature-scaled entropy weighted by to down-weight dubious, high-entropy samples. Incorporate prototype alignment, with weight 0.05, between the global mean of the last feature map and an online source prototype , which provides a lightweight semantic–geometric anchor that mitigates the classification confidence up, localization down effect. Finally, apply gradient regularization, , with clipping (cap 3.5).
Risk gate. Maintain an EMA baseline of the gradient norm. When the current exceeds the baseline, trigger the breaker and abort the update for that sample. When the loss change between consecutive steps falls below , apply early stopping. Throughout, only the BN-affine parameters are updated.
For consistency, is initialized at the start of the session using a brief clean segment and maintained thereafter with an EMA. If a clear domain shift is detected, the prototype is refreshed to prevent anchoring to stale semantics.
Path C (high-noise isolation, optional rollback). When , the safest write is no write. The sample is skipped directly and, by policy, BN rollback is performed at the end of the batch or session to fully erase any suspect writes. Whereas AdaBN would continue aligning statistics and Tent would continue reducing entropy on low-quality frames, this path quarantines high noise outside the learning loop. is chosen to balance skip rate against benefit. If it is set too low, useful frames are rejected. If it is set too high, risky frames are admitted. A practical starting point is 1.8, after which tuning can proceed using a validation “skip–gain” curve.
2.5 Outputs and long-term stability
All three paths output standard detections. Route shares (A/B/C) and threshold traces are recorded online for audit and ablation. BN rollback is performed on a per-batch/session basis to confine adaptation effects to a local time window and to prevent slow accumulated drift/forgetting. In continuous deployments, health checks (e.g., recent-window , , skip ratio, grad-norm anomaly rate) and self-healing actions (forced rollback, prototype re-estimation, temporary increase of ) are defined to close the loop.
3 Experiments
3.1 Datasets and experimental settings
In our experiments, we use the Road Signs public dataset from Roboflow Universe (released by Warsaw University of Technology) for road-traffic sign detection. The dataset provides bounding-box annotations and covers diverse lighting, weather, and roadway conditions, which makes it well suited for evaluating generalization and robustness in realistic settings. To ensure reproducibility, we rely exclusively on the publicly provided data and labels. When no official split exists, we construct a stratified 8/1/1 train/val/test partition at the image level with a fixed random seed. All images are resized to and normalized using the ImageNet mean and standard deviation before entering the network. During training, we apply lightweight augmentations: random horizontal flip, rotation, slight Gaussian blur, and brightness/contrast perturbations. Validation and testing use no augmentation to preserve fairness. We report mAP0.5 and mAP@[0.5:0.95], and we additionally record inference latency to assess deployability under edge constraints.
Experiments were run on a laptop with an Intel Core i7-12700H (2.30 GHz), 16 GB RAM, and NVIDIA GeForce RTX 3060 (6 GB). The software stack includes Python 3.8.20, PyTorch 2.4.1+cu118/torchvision 0.19.1+cu118/torchaudio 2.4.1+cu118 (CUDA 11.8, cuDNN 8.8), and core dependencies NumPy 1.24.3, OpenCV 4.12.0.88, Pillow 8.4.0, and SciPy 1.10.1. For reproducibility, we fix the random seed to 42 and enable deterministic execution (disable cuDNN benchmark, enable deterministic algorithms). The model is EfficientDet + AdaTent++. Only BN-affine parameters are updated (Adam, LR ). Inputs are resized to and normalized by ImageNet mean/std. Routing uses the entropy consistency gate with EMA thresholds and hysteresis defined as , , , and . The skip threshold is (slightly higher in adverse-weather/blur domains). Path A applies smoothed BN blending with . Path B performs minimal backprop on BN-affine only with a hybrid objective: temperature-scaled entropy + prototype alignment (MSE weight 0.05) + gradient regularization , plus gradient clipping and a grad-norm breaker . The step count was dynamic (1–2) and dropped to 1 when the deviation from thresholds exceeded 0.35.
We warm up BN statistics on 27 images. The default test-time adaptation used with domain-adaptive view selection. Detection confidence was 0.05 in adverse-weather domains and 0.02 otherwise. Evaluation uses visual object class (VOC) mAP (intersection over union (IoU) , score ) and logs A/B/C path shares to monitor adaptation behavior. The RTX 3060 laptop used in this study is not an embedded or in-vehicle computing platform. Therefore, the reported experiments should not be interpreted as direct validation of real-time embedded deployment. The present work primarily evaluates detection accuracy and adaptation behavior under a resource-conscious implementation setting. Complete system-level profiling, including end-to-end latency, FPS, peak memory consumption, energy consumption, and path-specific adaptation overhead, is beyond the scope of the current evaluation.
3.2 Evaluation criteria
We evaluate detection performance with a Pascal-VOC protocol at IoU = 0.5 (mAP0.5) on a fixed, deterministic subset of 1,290 test images per domain. To construct this set, we selected 30 representative frames for each of the 43 traffic sign classes, ensuring that the selected instances cover diverse lighting, physical scales, and scene conditions to provide a statistically reliable baseline before applying corruptions. For each domain, the prescribed corruption is applied, the model predicts after an internal resize, and all boxes (predicted and ground truth) are evaluated in the original image coordinates with coordinates clipped to image bounds. Post-processing is per class: we apply non-maximum suppression at IoU = 0.5, then remove predictions below a domain gate (default 0.02, and 0.05 for adverse-weather domains to suppress long-tail false positives). Ground truths follow the VOC convention. Objects marked difficult are retained but ignored for matching. Matching is greedy, one-to-one, in descending score order within each class: a prediction is a true positive iff it matches a yet-unmatched ground truth of the same class with IoU 0.5. Otherwise, it is a false positive. The remaining ground truths are false negatives. A score threshold of 0.5 is enforced when forming precision–recall curves. Average precision (AP) is computed as the area under each class’s precision–recall curve (interpolated), and mAP0.5 is the unweighted mean AP across classes, reported as a percentage.
3.3 Comparative experiments
Source-only (none): No adaptation—run the source-trained model as is; zero overhead but cannot correct domain shift.
AdaBN: Recompute BN running mean/variance to align statistics; no backprop; effective for mild covariate shift.
Tent: Minimize prediction entropy with a few gradient steps on BN affine; helpful but can amplify wrong confidence on noisy frames.
SHOT: Fix the classifier and adapt features via information maximization and prototype alignment; relies on stable high-confidence pseudo-labels.
T3A: Gradient-free update of class prototypes from confident target features; extremely fast but limited under severe corruptions.
CoTTA: Teacher–student EMA with strong/weak augmentations and stochastic restoration; suitable for continual streams but more computationally involved.
According to Table 1, AdaTent++ yields the highest reported macro-average mAP50 (68.87%), exceeding T3A (68.03%) by 0.84 percentage points in a deterministic single-run evaluation. Because repeated runs and confidence intervals are not available, this difference should be interpreted as a descriptive numerical advantage rather than evidence of statistical superiority. Per domain, it ranks first on T2_dark 77.35% (+12.35 pp over the next best), T8_motion_blur 71.06% (+5.03 pp), and T16_composite 77.17% (+0.90 pp), and remains close to the top under weather corruptions with T17_rain 73.73% and T19_snow 72.98%. In structurally destructive regimes, including T3_blur, T20_fog, and T21_lens_blur, specialized baselines lead (81.96%, 80.45%, and 74.08%, respectively), while AdaTent++ yields 58.18%, 70.40%, and 52.72%, yet still improves over the source-only model in most cases.
TABLE 1
| Domain | None | AdaBN | SHOT | Tent | T3A | CoTTA | AdaTent++ (ours) |
|---|---|---|---|---|---|---|---|
| T2_dark | 54.53 | 69.90 | 54.70 | 57.75 | 65.00 | 60.32 | 77.35 |
| T3_blur | 63.56 | 68.83 | 66.57 | 50.94 | 81.96 | 60.74 | 58.18 |
| T8_motion_blur | 56.27 | 57.21 | 62.96 | 62.97 | 66.03 | 63.16 | 71.06 |
| T11_saltpepper | 57.30 | 60.32 | 69.38 | 70.75 | 65.50 | 59.44 | 58.56 |
| T13_low_brightness | 58.53 | 74.57 | 65.21 | 66.39 | 65.30 | 84.16 | 72.83 |
| T14_low_contrast | 61.72 | 55.57 | 79.20 | 54.20 | 65.38 | 66.62 | 71.72 |
| T16_composite | 67.36 | 51.93 | 60.05 | 61.51 | 76.27 | 64.34 | 77.17 |
| T17_rain | 62.06 | 69.80 | 65.38 | 74.57 | 68.38 | 71.14 | 73.73 |
| T18_night | 62.94 | 58.27 | 76.55 | 50.37 | 62.91 | 58.29 | 69.76 |
| T19_snow | 58.56 | 64.15 | 68.80 | 66.82 | 75.48 | 65.84 | 72.98 |
| T20_fog | 71.94 | 68.05 | 69.72 | 80.45 | 65.13 | 54.71 | 70.40 |
| T21_lens_blur | 69.72 | 59.44 | 68.69 | 57.67 | 63.68 | 74.08 | 52.72 |
| Average | 62.04 | 63.17 | 67.27 | 62.87 | 68.03 | 65.24 | 68.87 |
Per-domain mAP comparison (%).
The best result in each row is in bold.
Qualitative evidence in Figures 2–4 corroborates these measurements: under dark/defocus (Figure 2), AdaTent++ produces stable, high-confidence boxes where competitors miss or jitter, consistent with its large margin on T2_dark and solid behavior when blur is motion-induced (T8_motion_blur). Under motion blur vs. salt-and-pepper noise (Figure 3), it localizes tightly in the former but is less effective in the latter, mirroring the drop on T11_saltpepper where fine structure is destroyed. Under low light and rain-like weather (Figure 4), it reduces false positives and maintains detections, aligning with near-top scores on T17_rain and T19_snow. AdaTent++ achieves the highest macro-average mAP50 in the reported experiment, but it is the best-performing method in only three of the twelve domains. It performs strongly under dark, motion blur, and composite shifts, while T3A, Tent, CoTTA, or SHOT are superior in several blur, noise, fog, and lens-blur conditions. The result therefore supports average robustness in the studied setting, not consistent per-domain superiority.
FIGURE 2
FIGURE 3
FIGURE 4
3.4 Ablation study
As shown in Table 2, we perform a leave-one-out ablation starting from the full AdaTent++ configuration (mAP@0.5 = 68.87, averaged over 12 domains). The largest degradations arise when disabling components that govern reliability: removing BN-state rollback reduces performance to 62.38 ( percentage points), removing routing reduces it to 62.71 ( pp), and disallowing the skip path for high-noise frames yields 62.78 ( pp). These declines indicate that rollback limits error accumulation across frames, while routing plus skip isolates unreliable inputs and prevents harmful updates. The gating rule itself is also consequential: replacing EMA thresholds with fixed cutoffs lowers mAP to 64.14 ( pp), and using entropy alone (without cross-view agreement) yields 64.42 ( pp), suggesting that adaptive thresholds track non-stationary conditions and agreement contributes complementary evidence beyond uncertainty magnitude. On the update objective and schedule, removing prototype alignment produces 64.83 ( pp), consistent with the value of a lightweight semantic–geometric anchor, whereas dropping gradient regularization gives 67.99 ( pp), a smaller but stabilizing effect. The step policy exhibits a bias–variance trade-off: a fixed single step underfits (63.49, pp), three fixed steps tend to over-update (67.33, pp), while the dynamic 1–2-step policy used in AdaTent++ achieves the best balance. Finally, the multi-view size shows an optimum at K = 2: K = 1 is slightly worse (68.25, pp), and K = 4 is detrimental (64.23, pp), plausibly due to noisier consensus and increased latency. Taken together, the evidence supports a joint mechanism: reliability-gated routing with adaptive thresholds and a skip path, safeguarded by BN rollback, combined with a minimal BN-affine update under a guarded hybrid objective and conservative scheduling (dynamic steps, small K) to control test-time variance. Table 2 is an aggregate leave-one-out ablation averaged over 12 domains. It estimates the first-order effect of removing one component from the implemented pipeline, but it does not isolate higher-order interactions, per-domain behavior, or statistical variability. The observed drops should therefore not be interpreted as independent causal contributions. Per-domain and factorial ablations remain necessary for a fuller analysis.
TABLE 2
| Variant | mAP | |
|---|---|---|
| AdaTent++ (full) | 68.87 | – |
| Ours, K = 1 | 68.25 | |
| Ours, no grad-reg | 67.99 | |
| Ours, fixstep = 3 | 67.33 | |
| Ours, no prototype align | 64.83 | |
| Ours, entropy only | 64.42 | |
| Ours, K = 4 | 64.23 | |
| Ours; fixed threshold | 64.14 | |
| Ours, fixstep = 1 | 63.49 | |
| Ours, no skip path C | 62.78 | |
| Ours, no routing | 62.71 | |
| Ours, no BN backup | 62.38 |
Ablation results (mAP %). is the drop vs. full AdaTent++.
4 Discussion
The evidence is consistent across three views: Table 1 (highest average mAP0.5), Figures 2–4 (qualitative trends matching the per-domain results), and the leave-one-out ablation (which identifies load-bearing components), and supports a single mechanism: a reliability-first adaptation policy that remains stable under test-time constraints. Entropy combined with cross-view agreement, moderated by EMA-plus-hysteresis thresholds, yields stream-aware, non-flapping routing. Path A (smoothed AdaBN) absorbs benign covariate drift without backprop, Path B (minimal BN-affine update) applies a guarded hybrid objective with dynamic 1–2 steps to cap variance and latency, Path C (skip/rollback) blocks harmful writes on high-noise frames, and BN-state rollback limits error accumulation over time. This architecture explains the strong margins where shift is chiefly covariate (darkness, motion blur, and composite) and the overall lead in average accuracy, while also clarifying expected shortfalls under structurally destructive corruptions (severe defocus, dense fog, and impulse noise) where entropy/agreement become unreliable and BN-level updates cannot reconstruct lost low-level structure. Three low-overhead extensions are warranted: augment gating with task-agnostic quality cues or box-level uncertainty, allow ultra-light low-level enhancement or shallow gated adaptation within the same latency envelope, and add detection-specific consistency (e.g., IoU-aware entropy, cross-view box reprojection). Although all baselines share the same source backbone, test image sequence, preprocessing, post-processing, and evaluation code, their native adaptation procedures entail different computational costs. The present results should therefore be interpreted as an accuracy-oriented comparison under a unified evaluation protocol rather than as a strictly compute-matched benchmark. Future evaluations should span longer streams, broader devices, and additional tasks. For deployment, windowed rollback, monitoring of threshold/routing traces, and anomaly triggers with fallback to the source model are practical safeguards against long-horizon drift. Overall, the gains arise from the combination of reliability gating, conservative scheduling, and BN-level minimal updates, delivering a measurable stability–efficiency–accuracy trade-off suitable for edge-constrained test-time adaptation.
While AdaTent++ achieves the highest average mAP, it predictably underperforms in structurally destructive domains (e.g., T3_blur, T11_saltpepper, T20_fog, or T21_lens_blur). In these scenarios, low-level image features are severely obliterated, causing prediction entropy and multi-view inconsistency to spike. Consequently, the dynamic routing mechanism categorizes most of these frames as highly unreliable, routing them to Path C (Skip Adaptation). While this conservative behavior effectively prevents the catastrophic drift observed in baseline methods (preserving overall average accuracy), it limits the model’s ability to adapt to severe structural loss, as minimal BN-affine updates lack the capacity to reconstruct missing spatial features. Future work will explore integrating task-agnostic quality cues (e.g., SNR and haze indices) to better calibrate the routing thresholds under these extreme corruptions.
Finally, while our experiments rigorously evaluate 12 corrupted domains, they are currently constrained to a single primary source dataset due to the scarcity of multi-source TSR datasets with aligned, fine-grained bounding-box annotations. A critical direction for our future work is the collection and annotation of a proprietary, large-scale TSR dataset to facilitate extensive cross-dataset validation and further confirm the method’s cross-regional generalization.
5 Limitations
The present study is restricted to one public dataset, a fixed class-balanced subset, 12 constructed corruption domains, and EfficientDet-D0. The evaluation is domain-aware at the configuration level because benchmark condition identifiers are used for view selection and candidate filtering. The reported 0.84-percentage-point improvement is based on a deterministic single run and has not been supported by repeated-run statistics or confidence intervals. The method uses a single-anchor reliability proxy and class-level cross-view agreement, without an explicit localization-consistency objective. In addition, the laptop-GPU evaluation does not establish embedded deployment, and no long continuous-stream experiment is provided. Future work will address full-test-set and cross-dataset evaluation, a globally fixed domain-agnostic protocol, object- and box-level reliability, additional detector architectures, repeated-run statistics, embedded profiling, and continual-stream stability.
6 Conclusion
We addressed test-time adaptation for object detection under edge constraints by proposing AdaTent++, a reliability-first framework that (i) gates updates using entropy cross-view agreement with EMA-plus-hysteresis thresholds and (ii) executes one of three mutually exclusive routes per frame: Path A (smoothed AdaBN, zero backprop) to absorb benign covariate drift, Path B (minimal BN-affine update) with a guarded hybrid objective and dynamic 1–2 steps, and Path C (skip/rollback) to prevent harmful writes on high-noise inputs. Under a shared backbone, identical preprocessing, and matched test-time budgets, AdaTent++ achieves the highest average mAP@0.5 (68.87%) over 12 corrupted domains (Table 1), with pronounced gains where shift is predominantly covariate (e.g., T2_dark 77.35%, T8_motion_blur 71.06%, and T16_composite 77.17%) and competitive performance in weather corruptions. A leave-one-out ablation confirms that reliability controls (routing, skip, and BN-state rollback) and adaptive thresholds are load-bearing, while conservative scheduling (small K and dynamic steps) and light regularization (prototype alignment and gradient control) further stabilize updates. Collectively, the results substantiate a consistent mechanism: BN-level minimal updates governed by a stream-aware reliability gate deliver a measurable stability–efficiency–accuracy trade-off suitable for edge deployment.
Building on these findings, we will (1) strengthen the reliability estimator with orthogonal, low-overhead signals: task-agnostic quality cues (sharpness, SNR, and haze indices) and box-level uncertainty calibration, to improve gating under structurally destructive corruptions. (2) Introduce ultra-light, gated low-level enhancement or shallow constrained adaptation (e.g., low-rank/depthwise layers) within the same latency envelope to restore edge/detail before BN-level adaptation. (3) Add detection-specific consistency objectives (e.g., IoU-aware entropy, cross-view reprojection agreement) to reduce classification–localization mismatch. (4) Develop budget-aware controllers that adapt K and step count online given latency/energy budgets while preserving stability guarantees. (5) Expand external validity via longer streams, broader devices, and additional tasks (segmentation and keypoints) under the same protocol, reporting accuracy–latency–energy jointly. (6) Pursue theoretical analysis of the gate’s decision stability (e.g., sensitivity bounds under distribution shift) and of rollback schedules over long horizons. These directions aim to generalize AdaTent++ beyond the current setting while retaining its minimal-update principle and deployment practicality.
Statements
Data availability statement
The data sets analyzed for this study can be found in the Road Signs public dataset from Roboflow Universe (released by Warsaw University of Technology) for road-traffic sign detection: https://universe.roboflow.com/politechnika-warszawska-yvotb/road-signs-o6iyq.
Author contributions
XG: Conceptualization, Writing – review and editing. MW: Conceptualization, Writing – review and editing; YG: Data curation, Writing – review and editing. SW: Formal Analysis, Writing – review and editing. YL: Writing – original draft, Writing – review and editing. RF: Writing – original draft, Writing – review and editing.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This work was supported in part by the Shandong Provincial Natural Science Foundation under Grant ZR2025QC649, in part by the China Postdoctoral Science Foundation under Grant 2024M752301, in part by the Weifang Soft Science under Grant 2024RKX074, and in part by the Weifang University of Science and Technology A-Class Doctoral Research Fund under Grant KJRC2024006.
Acknowledgments
The authors thank everyone for their help and support.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsip.2026.1842851/full#supplementary-material
References
1
FanZ.WanZ.LiuC.-K.LuA.BhardwajK.RaychowdhuryA. (2024). Benchmarking test-time dnn adaptation at edge with compute-in-memory. ACM J. Auton. Transp. Syst.1, 16–26. 10.1145/3665898
2
HuseljicD.HerdeM.HahnP.MüjdeM.SickB. (2025). Of uncertainty calibration in pretrained object detectors. Int. J. Comput. Vis.133, 1033–1047. 10.1007/s11263-024-02219-z
3
LeeH.LeeS.JungI.HongS. (2025). Prototypical class-wise test-time adaptation. Pattern Recogn. Lett.187, 49–55. 10.1016/j.patrec.2024.10.011
4
LiY.WangN.ShiJ.HouX.LiuJ. (2018). Adaptive batch normalization for practical domain adaptation. Pattern Recogn.80, 109–117. 10.1016/j.patcog.2018.03.005
5
LiuY.WangJ.HuangC.WuY.XuY.CaoX. (2024). Mlfa: toward realistic test time adaptive object detection by multi-level feature alignment. IEEE transactions on image processing :a publication of the IEEE Signal Processing Society33, 5837–5848. 10.1109/TIP.2024.3473532
6
NiuS.WuJ.ZhangY.ChenY.ZhengS.ZhaoP.et al (2022). Efficient test-time Model Adaptation Without Forgetting.
7
NiuS.WuJ.ZhangY.WenZ.ChenY.ZhaoP.et al (2023). Towards Stable test-time Adaptation in Dynamic Wild World.
8
QuanZ.SunJ. (2025). A feature-enhanced small object detection algorithm based on attention mechanism. Sensors25, 589. 10.3390/s25020589
9
RafiT. H.MahjabinR.GhoshE.KoY.-W.LeeJ.-G. (2024). Domain generalization for semantic segmentation: a survey. Artif. Intell. Rev.57, 247. 10.1007/s10462-024-10817-z
10
RivoirD.FunkeI.SpeidelS. (2024). On the pitfalls of batch normalization for end-to-end video learning: a study on surgical workflow analysis. Med. Image Anal.94, 103126. 10.1016/j.media.2024.103126
11
TanM.ChenG.WuJ.ZhangY.ChenY.ZhaoP.et al (2025). Uncertainty-calibrated test-time model adaptation without forgetting. IEEE Transactions Pattern Analysis Machine Intelligence47, 6274–6289. 10.1109/TPAMI.2025.3560696
12
TianL.OhC.CavallaroA. (2024). Test-time adaptation for 6d pose tracking. Pattern Recogn.152, 110390. 10.1016/j.patcog.2024.110390
13
WangD.ShelhamerE.LiuS.OlshausenB.DarrellT. (2021). Tent: Fully test-time Adaptation by Entropy Minimization.
14
WangQ.FinkO.GoolL. V.DaiD. (2022). Continual test-time Domain Adaptation.
15
WangZ.LuoY.ZhengL.ChenZ.WangS.HuangZ. (2025). In search of lost online test-time adaptation: a survey. Int. J. Comput. Vis.133, 1106–1139. 10.1007/s11263-024-02213-5
16
YangH.WangM.YuZ.ZhangH.JiangJ.ZhouY. (2024). Confidence-based and sample-reweighted test-time adaptation. Knowl.-Based Syst.283, 111164. 10.1016/j.knosys.2023.111164
Summary
Keywords
adaptive batch normalization, domain generalization, dynamic routing, EfficientDet, entropy minimization, object detection, test-time adaptation
Citation
Gai X, Wang M, Guo Y, Wang S, Li Y and Fu R (2026) AdaTent++: reliability-guided hybrid test-time adaptation for traffic sign detection under distribution shift. Front. Signal Process. 6:1842851. doi: 10.3389/frsip.2026.1842851
Received
31 March 2026
Revised
31 July 2026
Accepted
03 August 2026
Published
09 September 2026
Volume
6 - 2026
Edited by
Maryam Ravan, New York Institute of Technology, United States
Reviewed by
Lesia Mochurad, Department of Artificial Intelligence, Ukraine
Xiaoqian Ruan, University of Illinois Chicago, United States
Updates
Copyright
© 2026 Gai, Wang, Guo, Wang, Li and Fu.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Rui Fu, furui19891209@wfust.edu.cn
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.