PaperlayerView exact revision on arXiv

Source-linked AI summary

Evidential Deep Learning for Multi-Modal Anti-UAV Detection

Dmitry Golovchits, Seyed Sahand Mohammadi Ziabari, Ali Mohammed Mansoor Alsahag

arXiv:2609.01742v1cs.CV

TL;DR

Anti-UAV systems lack reliable per-modality reliability signals despite increasingly multimodal sensing. This study uses controlled three-benchmark ablations to separate evidential training, DS fusion, and uncertainty-driven temporal gating. Evidential training improves predictive performance, but DS fusion and uncertainty-based control do not deliver their intended advantages.

  • Problem

    Anti-UAV multimodal systems lack explicit per-modality reliability signals, leaving unresolved whether evidential objectives, evidence fusion, and uncertainty can improve detection and sensor management.

  • Method

    EviDS-UAV evaluates evidential detection heads, DS evidence fusion, and temporal sensor gating independently across three complementary anti-UAV benchmarks.

  • Results

    The evidential training objective improves accuracy, while DS fusion does not beat averaging, uncertainty fails for detection-level error ranking, and temporal gating provides no realized latency gain.

  • Takeaways & Limitations

    The main benefit of evidential learning comes from its training objective rather than its uncertainty estimate, with gains not extending to fusion or efficient sensor selection.

  • Takeaways & Limitations

    The evaluation uses 50 held-out validation sequences rather than the unavailable competition test split, and E3 tracking reports a single seed.

Abstract

from arXiv · show

Anti-UAV systems increasingly fuse multiple sensors, yet their detection heads provide no per-modality reliability signal. This study evaluates whether evidential deep learning (EDL) heads, Dempster-Shafer (DS) evidence fusion, and uncertainty-driven temporal sensor gating improve anti-UAV detection through a controlled ablation on three benchmarks: thermal tracking (AntiUAV600), RGB-audio-RF classification (TRIDENT), and RGB-IR tracking (MM-UAV). The EDL training objective improves accuracy over retrained sigmoid baselines (+5.9 percentage points in accuracy and a tripled tracker-on-absent rate in E1; +4.8 percentage points in classification accuracy in E2, surviving a clip-clustered bootstrap, p = 0.011) and ranks classification errors substantially better (entropy UAUC approximately 0.94 vs. 0.51). The remaining components do not support their respective hypotheses. DS fusion does not outperform simple probability averaging. Dirichlet vacuity adds no ranking power beyond predictive entropy and inverts at the detection level, where extreme background imbalance causes it to encode class membership rather than error likelihood, a failure also observed for entropy and sigmoid confidence. Temporal gating preserves accuracy only when nearly inactive and yields no realised latency saving on shared-backbone hardware. The benefit of evidential learning therefore arises primarily from its training objective rather than its uncertainty estimate; a crop-level control further localises the detection-level breakdown to anchor-level evaluation rather than the learned representation.

1 Introduction

The study addresses the need for reliable anti-UAV detection across degraded, multimodal sensor conditions by testing evidential heads, DS fusion, and uncertainty-driven gating. Controlled ablations separate the contribution of each component rather than assuming that the combined design improves performance.

  • Anti-UAV systems must detect, classify, and track reliably across diverse and degraded sensor conditions.
  • Standard confidence scores do not distinguish evidence-backed predictions from outputs caused by sensor noise or out-of-distribution inputs.
  • Multi-modal systems commonly use learned gating or attention without an explicit per-sensor reliability model.
  • EviDS-UAV tests evidential Dirichlet outputs, Dempster-Shafer fusion, and uncertainty-driven modality gating as modular replacements for conventional components.
  • Controlled ablations show that evidential training improves accuracy and detection-tracking switching, whereas DS fusion, vacuity filtering, and temporal gating do not outperform simpler alternatives.

2 Related Work

Prior work separately advances anti-UAV perception, multimodal fusion, and evidential uncertainty, but the generality and practical value of combining them remain unresolved. This study targets that gap across heterogeneous sensor combinations and task levels.

  • Anti-UAV detection and tracking: Anti-UAV pipelines progressed from single-modality detection toward tracking, but standard detector confidence does not reliably propagate downstream.
  • Multimodal fusion: Multimodal fusion reduces sensor blind spots, while benchmarks expose persistent cross-modal misalignment and modality-specific degradation.
  • Evidential uncertainty: Bayesian approximations, ensembles, calibration, and EDL provide alternative uncertainty estimates, with DS theory enabling evidence-based fusion.
  • Research gap: The unresolved questions concern whether evidential objectives help detection independently, whether evidence fusion beats averaging, and whether uncertainty can drive temporal gating.

3 Methodology

EviDS-UAV adds evidential heads, DS fusion, and temporal gating to existing baselines without changing feature extraction or tracking modules. The method derives probabilities and uncertainty from Dirichlet evidence, fuses modality evidence, and evaluates gating under controlled ablations.

  • Framework: EviDS-UAV is a modular three-component method applied without modifying existing feature extraction or tracking modules.
  • Evidential detection heads: Standard sigmoid or softmax outputs are replaced by a Dirichlet evidential head that converts nonnegative evidence into class probabilities and epistemic uncertainty.
  • Evidential detection heads: High total evidence drives uncertainty toward zero, whereas ambiguous or out-of-distribution inputs retain low evidence and uncertainty near one.
  • Dempster-Shafer fusion: DS fusion combines modality-specific belief masses and uncertainty, weighting contributions by evidence strength and applying the rule iteratively for more than two modalities.
  • Dempster-Shafer fusion: DS fusion assumes independent evidence sources and becomes unstable near complete inter-modal conflict, requiring empirical evaluation and a fallback for high conflict.
  • Temporal gating: Hard temporal gating disables a modality after its uncertainty exceeds a threshold, using hysteresis, a suppression window, and an at-least-one-active-modality constraint.
  • Training and evaluation: The EDL loss replaces only the classification loss, retaining baseline regression and auxiliary losses while adding annealed KL regularization.
  • Training and evaluation: Retrained corrected baselines provide fair ablation comparators, while E1 uses TrkOnAbs as its primary switching criterion and the 50-sequence validation set is not leaderboard-comparable.

4 Results

Across three benchmarks, the evidential training objective improves predictive performance, while DS fusion and uncertainty-driven gating provide no consistent practical advantage. Uncertainty ranking is task-level dependent: useful for balanced classification, but inverted or ineffective at detection level.

  • E1: Evidential Detection: +5.9 pp accuracy and a tripled tracker-on-absent rate are achieved by EDL-270 over Sigmoid-270 in E1.The higher TrkOnAbs indicates fewer false-positive tracks and more reliable switching on background frames.
  • E1: Evidential Detection: Detector vacuity inverts at detection level, reaching 0.44 on correct detections versus 0.019 on errors and reducing both Acc and TrkOnAbs when thresholded.The failure suppresses correct re-acquisitions and creates extended detector-mode streaks.
  • E2: TRIDENT: DS fusion does not significantly outperform averaging: DS reaches 97.53% versus 97.20%, with a +0.33 pp difference and p = 0.332.Under full degradation, averaging beats DS on accuracy, macro-F1, and ECE; the practical equivalence is robust to belief-mass construction.
  • E2: TRIDENT: EDL entropy ranks classification errors substantially better than the sigmoid baseline, with UAUC 0.943 for DS and 0.942 for Average versus 0.510 for Late Fusion.Predictive entropy matches or exceeds vacuity, so vacuity adds no usable ranking power beyond the Dirichlet mean.
  • Gating: Temporal gating produces a 27–32% mean skip rate in E2 but no operationally useful trade-off, while detection-level gating in E3 never skips a stream and exactly reduces to DS.In E3, bimodal vacuity provides no usable middle range for sustaining the gate-off condition.
  • E3: MM-UAV: In E3, evidential heads collapse tracking at the published threshold, but decision-level fusion restores tracking partially: averaging reaches RGB MOTA 57.12 and DS 55.01.The evidential-head collapse is a confidence-scale artefact; fusion before thresholding lifts scores above 0.3 for detections supported by either stream.
  • E3: MM-UAV: Detection-level uncertainty is anti-correlated with error across conditions, with vacuity-UAUC 0.03–0.15 and predictive-entropy-UAUC 0.04–0.10 for evidential conditions.The lowest uncertainty occurs on confident false positives, so detector-level uncertainty gating is unsupported in E3.

5 Discussion

Across the ablations, evidential training improves predictive behavior, while DS fusion and uncertainty-driven gating fail to outperform simpler alternatives. Uncertainty is useful for balanced classification but breaks down at dense detection level, with the failure localized to anchor-level evaluation.

  • DS fusion provides no advantage over probability averaging in E2 and performs worse in E3.The added uncertainty-aware fusion mechanism does not improve predictive performance under the evaluated conditions.
  • EDL improves E1 accuracy and approximately triples the tracker-on-absent rate, but vacuity does not reliably discriminate detector errors.The controlled ablation separates the optimization effect of the Dirichlet objective from the inference-time value of vacuity.
  • Uncertainty-driven gating provides no meaningful efficiency benefit and degrades accuracy beyond a narrow low-skip regime.Shared-backbone hardware still executes all modality encoders before gating, while E3 lacks sufficient vacuity dynamic range for threshold selection.
  • 5.1 Evidential Uncertainty at the Detection versus Classification Level: Predictive uncertainty separates errors in balanced classification but breaks down under extreme class imbalance at detection level.E2 reaches UAUC up to 0.94, whereas dense detection signals collapse or invert when background anchors dominate.
  • 5.1 Evidential Uncertainty at the Detection versus Classification Level: A crop-level control raises vacuity UAUC from 0.036 at anchor level to 0.628, localizing the inversion to dense anchor evaluation rather than learned representation.The same frozen weights recover usable error ranking when vacuity is measured over isolated detection crops.
  • 5.2 Why Dempster-Shafer Fusion’s Advantages Did Not Translate: DS fusion’s theoretical advantage requires conflict that distinguishes reliable from unreliable modalities; absent that signal, averaging matches or outperforms it.In E3, the focal-class setting makes conflict structurally zero, while decision-level fusion also incurs localization and duplication penalties.
  • 5.3 Practical Implications: For the studied redundant, imbalanced systems, evidential training is the preferred drop-in improvement, while averaging is the lower-complexity default for fusion.The paper cautions against using vacuity, entropy, or sigmoid confidence for gating or false-positive filtering without target-distribution validation.
  • 5.5 Validity and Reliability: The quantitative margins are setting-specific because E1 uses 50 validation sequences, E3 has a single seed, and effective sample size is bounded by clips or sequences.The E2 clip-clustered bootstrap uses n = 33 and widens confidence intervals three- to five-fold relative to naive segment-level intervals.

6 Conclusion

Across three benchmarks, the evidential training objective improves detection accuracy, while DS fusion, uncertainty estimates, and temporal gating do not deliver their intended benefits. The study therefore separates gains from the training objective from failures of the packaged uncertainty and control mechanisms.

  • Evidential training improves detection accuracy, but DS fusion, uncertainty-based control, and temporal gating do not improve their respective objectives.The asymmetric pattern holds across the controlled ablations: DS fusion does not improve robustness over averaging, uncertainty fails as a reliable control signal, and gating provides no efficiency gain.
  • DS fusion does not significantly outperform probability averaging in the evaluated settings.The comparison tests whether evidence-based fusion provides an advantage over a simpler multimodal fusion rule.
  • 5.9 percentage points in E1 and 4.8 percentage points in E2: replacing the classification head and loss improves accuracy over retrained sigmoid baselines.E1 also approximately triples the tracker-on-absent rate, while the E2 gain remains robust under clip-clustered bootstrap analysis.
  • Predictive uncertainty ranks classification errors better, but detection-anchor uncertainty becomes inversely associated with error under extreme background imbalance.A crop-level control recovers meaningful uncertainty from isolated small-object crops, localising the failure primarily to anchor-level measurement rather than the learned representation.
  • Per-modality uncertainty preserves E2 accuracy only when gating is nearly inactive, reduces accuracy by 5.7 percentage points at the selected operating point, and yields no E3 latency saving.On shared-backbone hardware, modality encoders execute before the gating decision; the E3 gate did not engage.
  • The quantitative margins are setting-specific because degradation is synthetic, E1 uses validation sequences, E2 has 33 independent clips, and E3 uses one tracking seed with two-modality fusion.The authors state that qualitative conclusions recur across the experiments despite these scope limitations.

A Training and Inference Protocols

The protocols preserve comparable baseline structures while applying evidential changes at specified heads or fusion stages. Training and inference choices differ across E1, E2, and E3, including frozen modules, segment-level evaluation, and a two-stage tracking pipeline.

  • Baseline selection: Baseline selection leaves detection heads with standard sigmoid outputs, confining evidential reasoning, where present, to downstream modules.This establishes the structural gap targeted by EviDS-UAV.
  • E1 training and inference protocol: E1 retrains only the YOLOv5s detector while retaining the released Stage-2 CvT tracker checkpoint.The protocol isolates the detector-side change from tracker retraining.
  • E2 training and inference protocol: E2 trains unimodal models independently, then trains only the fusion head with frozen backbones at 0.25-second segment intervals.The protocol separates modality-specific representation learning from fusion training.
  • E3 training and inference protocol: E3 pretrains the dual-stream YOLOX-s backbone, fine-tunes detection heads with alignment components as specified, and uses parameter-free DS or averaging fusion.Inference runs through MMA-SORT with event-based identity association.
  • TRIDENT unimodal architectures: E2 trains each modality independently, freezes backbone weights during fusion training, and uses an RF architecture with temporal depth 1 per 0.25-second segment.The RF branch is architecturally 3D but operates on one frame per segment.
  • Implementation details: The reported implementation includes temporal depth 1, reducing the 3D architecture to 2D convolution in practice, and omits the RF MobileNet variant.Five unimodal models were trained rather than six.

B Ablation Configurations

The ablation configurations use fixed splits, checkpoints, backbones, and evaluation protocols to isolate component effects across E1, E2, and E3. The designs include controlled noise and a common dual-stream tracking setup where applicable.

  • E1 ablation configurations: E1 ablations share a 270/30/50 data split, the same CvT tracker checkpoint, and an identical evaluation protocol.These controls keep data, tracker state, and evaluation fixed across conditions.
  • E2 ablation configurations: E2 ablations use the corrected codebase, frozen unimodal backbones, and identical noise protocols, with RF evaluated clean-only except where stated.Conditions d′ and g form a controlled foil pair differing only in fusion rule under matched per-modality noise.
  • E3 ablation configurations: E3 conditions share a 1080/120/121 train/tune/test split, dual-stream YOLOX backbone, OGAA alignment, and identical MOTA/HOTA/IDF1/IDs plus detection-ECE evaluation.Stage 2 is trained with backbone and FPN frozen for the specified conditions.

C Full Results Tables

The appendices provide complete benchmark results, reproduction notes, and metric caveats for E1, E2, and E3. Several reported values require careful interpretation because of prediction collapse, coordinate artifacts, or corrected baseline discrepancies.

  • Table C.1 covers 50 AntiUAV600 validation sequences, including 56,301 frames and 1,735 absent frames.
  • TrkOnAbs measures the fraction of absent frames where tracking remains active without delegating to the detector.
  • E2’s ResNet-10 (Video) UAUC is inflated because the model predicts Drone 0% of the time, collapsing to the minority class.
  • Condition (g) collapses to majority-class prediction, producing Drone predictions on 99.7–99.8% of segments and an artefactual UAUC.
  • For conditions (d), (e), and (f), per-stream IR metrics are structurally undefined because fused detections are returned in RGB coordinates.

E.1 E2 Robustness and Calibration Checks

Robustness checks show that DS-versus-average conclusions depend somewhat on belief-mass construction but remain statistically equivalent, while gating can damage clean accuracy and does not realise projected hardware savings.

  • Calibration: The evidential head does not improve detector calibration: detector ECE is essentially unchanged, while temperature scaling lowers it through monotonic confidence compression.
  • Gating efficiency: The gating benchmark always runs all encoders, so projected encoder savings of −1.64 ms are unrealised on the monolithic GPU.The measured gate-decision overhead is +0.26 ms; true skipping requires per-modality hardware or asynchronous dispatch.
  • Clean evaluation: 97.35% versus 99.02% clean accuracy shows that gating can harm performance when it fires on clean inputs under the canonical belief-mass construction.The gate fires on 11.7% of audio segments; under the alternative construction, it never flipped a decision.
  • Belief-mass construction: The alternative belief-mass construction reverses the DS-versus-average ordering, but the margin remains within bootstrap noise: 95% CI [−1.82, +3.03] pp and p = 0.32.Under the canonical construction, DS leads by 0.15–0.61 pp and the comparison has p = 0.332.

E.2 Crop-Level Control (E3)

The E3 crop-level control evaluates uncertainty on detection crops using the same frozen evidential model, while scale analysis indicates that aggregate recovery is dominated by small objects.

  • Crop-level control: The crop-level control uses 1.5× context detection crops resized to 640 × 640 with aspect-preserving zero-padding, without retraining.Crops come from analysis-floor detections on a 15% random subsample of test-set image IDs within the 121 test sequences.
  • Scale strata: Aggregate recovery for averaged evidence is carried by the small-object stratum, which dominates the benchmark.Small objects are defined as source-box areas below 200 px^2; medium and large strata are descriptive because their sample sizes are too small for clustered inference.

E.3 Uncertainty-Ranking Diagnostics (E1, E3)

Uncertainty-ranking diagnostics distinguish classification errors more effectively than detection errors, while temporal-gating sweeps expose a narrow accuracy–skip-rate trade-off and calibration asymmetry across task levels.

  • E1 diagnostics: At the detection level, vacuity is not reliably estimable in E1 because only 5–6 false positives occur across all sequences.
  • E1 diagnostics: E1 detector-vacuity effects across challenge attributes are negligible, with |d| < 0.15, and can point opposite to the expected difficulty signal.
  • Temporal gating: The temporal-gate sweep is bimodal: only low-skip configurations preserve accuracy, while projected latency decreases with skip rate.The sweep uses 1,320 test segments and 664 configurations over cached encoder outputs.
  • Calibration: Reliability diagrams show a near-diagonal tracker but a strongly miscalibrated detector for several E1 conditions, with temperature scaling improving detector calibration.

F Methodology Overview

Figure F.1 overviews the general evidential pipeline and the three experiments testing its individual components.

  • Figure F.1 presents the general evidential pipeline and three experiments that test individual components.
Loading 2609.01742v1…