PaperlayerView exact revision on arXiv

Source-linked AI summary

GEAR: From Dynamic Encoding to Dynamic Activation in Social Trajectory Prediction

Jiaheng Chen, Jiaxing Li, Leixia Wang, Jianan Ju, Tinghe Zhang

arXiv:2609.13778v1cs.CVcs.AI

TL;DR

Trajectory prediction methods often focus on encoding social context, with less attention to how it is activated during future generation. GEAR applies step-wise gates to individual-motion and social-resonance biases before trajectory composition. Experiments report improvements over the resonance-based baseline and competitive state-of-the-art performance across the evaluated datasets.

  • Problem

    Existing approaches give less explicit attention to how encoded social context should be activated during generation, despite interaction cues varying in strength across scenes and prediction steps.

    Source: §I · p. 2, B. Social Interaction Modeling and… · p. 2

  • Method

    GEAR uses step-wise activation gates to modulate self-motion and social-resonance biases before final trajectory composition.

    Source: §I · p. 2, B. Bias-Decomposed Generation and Its… · p. 4

  • Results

    On ETH-UCY, SDD, and NBA, GEAR achieves state-of-the-art or competitive performance and consistently improves the resonance-based baseline.

    Source: §I · p. 2

  • Takeaways & Limitations

    The findings support calibrating when social context is strengthened or suppressed during trajectory generation, alongside encoding it.

    Source: §V · p. 9

    Fixed composition assumes both bias terms remain fully active at every future step, an assumption that may not suit varying interaction contexts.

    Source: B. Bias-Decomposed Generation and Its… · pp. 3–4

Abstract

from arXiv · show

Human trajectory prediction requires modeling both individual motion patterns and social interactions among agents. Existing methods have made substantial progress by using attention mechanisms, graph structures, and temporal encoders to capture dynamic social context. However, most of them primarily focus on how social information is encoded, while paying less explicit attention to how the encoded social context should take effect during future trajectory generation. In this paper, we argue that dynamic social encoding does not necessarily imply dynamic social activation. The same interaction context may require different activation strengths across future horizons and scene densities: social cues should be strengthened when interaction evidence is strong, but suppressed when they are weak or noisy. To address this issue, we propose GEAR, a generation-aware bias activation model for human trajectory prediction. Built upon a bias-decomposed trajectory generation formulation, GEAR dynamically activates the individual-motion and social-resonance bias terms at each future step before final trajectory composition. This allows the model to explicitly control when and how strongly individual and social bias components participate in generation. Experiments on ETH-UCY, SDD, and NBA show that GEAR consistently improves the resonance-based baseline and achieves competitive state-of-the-art performance. Further analyses of activation patterns and density-grouped errors validate the importance of calibrating encoded social context during trajectory generation. Our code is available at https://github.com/11isnotavailable/GEAR.git.

I. INTRODUCTION

GEAR addresses the gap between dynamically encoding social context and dynamically activating it during trajectory generation. It uses step-wise bias gates and is evaluated on ETH-UCY, SDD, and NBA.

I. INTRODUCTION

  • GEAR targets trajectory prediction, where forecasts must account for both individual motion patterns and interactions with nearby agents.The introduction illustrates this contrast with straight motion in sparse scenes versus deceleration or yielding around peers.Source: §I · p. 1
  • Existing methods focus more on encoding social context than on activating it during generation, although fixed composition can over-activate weak cues or under-activate essential social constraints.Individual motion tends to dominate in short-term predictions and sparse scenes, while dense scenes call for stronger social activation.Source: §I · pp. 1–2
  • GEAR applies step-wise gates to decoded self-motion and social-resonance biases before final trajectory composition.Its gates use summaries of ego-motion deviation and ego-neighbor resonance relations to control when and how strongly each bias contributes.Source: §I · p. 2
  • On ETH-UCY, SDD, and NBA, GEAR achieves state-of-the-art or competitive performance, with activation patterns varying across prediction steps and scene densities.The analyses also report that GEAR consistently improves the resonance-based baseline.Source: §I · p. 2

II. RELATED WORK

Related work spans trajectory-prediction architectures, social-interaction modeling, and structured generation. GEAR focuses on calibrating decoded motion components during generation rather than only encoding or fusing information.

A. Human Trajectory Prediction

  • Trajectory prediction methods include recurrent, generative, waypoint-conditioned, memory-retrieval, graph, attention, Transformer, frequency-domain, diffusion, and structured-basis approaches.Many encode individual motion and social context before generating future positions.Source: A. Human Trajectory Prediction · p. 2

B. Social Interaction Modeling and Structured Trajectory Generation

  • Structured generation separates motion into components such as base trends, individual deviations, and interaction corrections, but often combines decoded components with a fixed rule.This can overlook the need to trade off motion continuity and social constraints across horizons and interaction intensities.Source: B. Social Interaction Modeling and… · pp. 2–3

C. Gating, Fusion, and Generation-Stage Modulation

  • Unlike encoding-stage gating or intermediate feature fusion, GEAR modulates decoded bias components at each future step before final trajectory composition.The activation adjusts component contributions according to prediction horizons and scene interaction intensity.Source: C. Gating, Fusion, and… · p. 3

III. METHODOLOGY

GEAR distinguishes decoding motion components from deciding how strongly they contribute during generation. Its prediction setup uses observed target and neighbor trajectories to forecast multiple candidate futures.

III. METHODOLOGY

  • GEAR learns step-wise gates to control the contributions of decoded self-motion and social-resonance biases to future generation.The approach focuses on final-stage activation rather than redesigning the social interaction encoder.Source: §III · p. 3, A. Problem Formulation · p. 3

A. Problem Formulation

  • The prediction task uses the target agent’s observed trajectory and neighboring agents’ observed trajectories to forecast the target’s future positions.The model generates K candidate future trajectories under the stated multimodal prediction protocol.Source: A. Problem Formulation · p. 3
  • Evaluation follows the best-of-K protocol, reporting minADEK and minFDEK with K = 20.This setting aligns with the resonance-based backbone and prior trajectory-prediction methods.Source: A. Problem Formulation · p. 3

B. Bias-Decomposed Generation and Its Limitation

The section represents future trajectories as a linear motion base plus self-motion and social-resonance biases. It motivates replacing fixed bias composition with adaptive activation before final composition.

B. Bias-Decomposed Generation and Its Limitation

  • The trajectory decomposition combines a linear motion base, a self-motion bias, and a social-resonance bias.The base extrapolates the target’s recent motion; the self-motion bias captures individual nonlinear deviation, and the social-resonance bias captures interaction-induced corrections.Source: B. Bias-Decomposed Generation and Its… · p. 3
  • Static composition assumes both bias terms are fully active at every future step.The passage states this assumption as gself(t) = 1 and gsoc(t) = 1 for all prediction steps.Source: B. Bias-Decomposed Generation and Its… · p. 3
  • Short-term or sparse-scene predictions may be harmed by excessive social correction, while dense scenes can benefit from social-resonance constraints.The proposed adaptation strengthens social correction when interactions are strong and suppresses it when interaction cues are weak.Source: B. Bias-Decomposed Generation and Its… · pp. 3–4
  • GEAR dynamically modulates decoded bias contributions before final trajectory composition rather than treating them as automatically active.The framework describes step-wise bias modulation within a bias-decomposed trajectory generator.Source: B. Bias-Decomposed Generation and Its… · p. 4, Figure 1 · p. 4

C. Generation-Aware Bias Activation

GEAR reuses the backbone’s decoded self-motion and social-resonance biases, then applies step-wise gates before final trajectory composition. The gates are generated from ego-motion deviation and ego-neighbor resonance summaries, preserving separate bias branches while allowing their contributions to vary over future steps.

C. Generation-Aware Bias Activation

  • GEAR modulates decoded self-motion and social-resonance biases at the final generation stage, rather than fusing features before decoding.The mechanism leaves bias generation and multimodal sampling unchanged.Source: C. Generation-Aware Bias Activation · p. 4
  • Mean-pooled ego-motion deviation and social-resonance summaries are concatenated and passed through a two-layer MLP with sigmoid gates.The gates are generated from fdiff, the ego trajectory deviation feature, and R, the ego-neighbor resonance relation matrix.Source: C. Generation-Aware Bias Activation · p. 4
  • The sigmoid-constrained gates separately control self-motion and social-resonance bias strength at each future step.Each gate is a length-Tpred sequence with values in [0, 1].Source: C. Generation-Aware Bias Activation · p. 4
  • The same gate sequences are shared across all K sampled predictions, so activation does not change candidate sampling.The gates derive from deterministic ego-motion and resonance summaries rather than sample-specific noise.Source: C. Generation-Aware Bias Activation · p. 4
  • Unlike static bias superposition, GEAR preserves separate linear, self-motion, and social-resonance components while varying their relative contributions over future horizons.Activation occurs after decoding and before final composition.Source: C. Generation-Aware Bias Activation · p. 4

D. Training Objective

GEAR is trained end-to-end using the resonance-based backbone’s protocol and a best-of-K trajectory regression objective.

D. Training Objective

  • GEAR uses a best-of-K trajectory regression objective given the ground-truth future and K predicted trajectories.Training follows the backbone protocol.Source: D. Training Objective · p. 5
  • The model outputs candidate future coordinates, with each trajectory represented as a linear base plus activated residual bias terms.It uses the same sampling and optimization protocol as the resonance-based backbone.Source: D. Training Objective · p. 5

IV. EXPERIMENTS

The experiments evaluate GEAR on standard trajectory-prediction benchmarks through comparisons with existing methods, ablations, and activation analyses.

IV. EXPERIMENTS

  • The experiments evaluate GEAR on standard trajectory-prediction benchmarks.Source: §IV · p. 5
  • The evaluation includes comparisons between GEAR and state-of-the-art methods.Source: §IV · p. 5
  • The experiments also include ablation studies and analyses of activation.Source: §IV · p. 5

A. Experimental Setup

GEAR is evaluated on ETH-UCY, SDD, and NBA using the Resonance protocols and improves that baseline across the reported benchmarks. Its applicability is more limited when another framework’s decoded components do not clearly correspond to personal and social motion.

A. Experimental Setup

  • The evaluation uses ETH-UCY’s five-scene leave-one-out protocol, SDD campus trajectories, and NBA game trajectories at tf = 5 and tf = 10.Reported metrics are best-of-20 minADE20/minFDE20, abbreviated ADE/FDE; lower values are better.Source: A. Experimental Setup · p. 5
  • GEAR retains Resonance’s linear-base, self-bias, and social-resonance branches, reusing fdiff and R to generate step-wise gates for final composition.The gate-generation process does not alter multimodal sampling or the number of candidates.Source: A. Experimental Setup · p. 5

B. Comparison with State-of-the-Art Methods

  • 0.16 versus 0.17 average ADE, with average FDE unchanged at 0.28: GEAR improves on the Resonance baseline on ETH-UCY.The reported results are competitive across all five leave-one-out scenes.Source: B. Comparison with State-of-the-Art… · p. 5, Table I · p. 6
  • 6.18/9.89 versus 6.27/10.02 ADE/FDE: GEAR achieves the best compared-method performance on SDD.The table reports ADE/FDE in pixels under best-of-20 evaluation; lower is better.Source: B. Comparison with State-of-the-Art… · p. 5, Table I · p. 6
  • For NBA at tf = 5, GEAR reduces FDE from 0.78 to 0.77 while maintaining ADE of 0.60; at tf = 10, ADE/FDE improve from 1.12/1.38 to 1.11/1.37.The reported gains occur under both short- and longer-horizon settings.Source: B. Comparison with State-of-the-Art… · p. 6, Table II · p. 6
  • On NSP, ADE remains 6.52 and FDE decreases from 10.52 to 10.51; the authors attribute the limited gains to components less explicitly aligned with personal and social motion.The authors note that methods without explicit decomposable motion components may not fully realize GEAR’s advantages.Source: B. Comparison with State-of-the-Art… · p. 6
  • Results across ETH-UCY, SDD, and NBA show improvements in pedestrian crowds, heterogeneous campus scenes, and sports scenarios.The authors interpret the cross-dataset consistency as supporting dynamic calibration during generation rather than fixed composition.Source: B. Comparison with State-of-the-Art… · p. 6

C. Ablation Study

The ablations support step-wise activation of both bias components: a social-only gate improves on Resonance, while full GEAR performs best. Comparisons with heavier variants and model-complexity results further characterize the design.

C. Ablation Study

  • Adding only social-bias activation improves on Resonance, and full GEAR achieves the best performance when it also activates the self-motion bias.The ablation supports coordinated activation of individual-motion and social-resonance biases.Source: C. Ablation Study · pp. 6–7, Table III · p. 7
  • GEAR’s self-motion and social-resonance activations vary over time, with stronger social activation in dense scenes.Fig. 2 visualizes the learned activation patterns.Source: Figure 2 · p. 7
  • Compared with Resonance, GEAR’s error reductions are larger in sparse scenes and increase toward longer prediction horizons.Fig. 3 reports step-wise gains for sparse and dense scenes.Source: Figure 3 · p. 7
  • Both attention-based fusion and mean+max pooling perform worse than GEAR, supporting decomposed bias roles with lightweight step-wise activation.The comparison suggests gains do not come simply from increasing fusion complexity.Source: C. Ablation Study · p. 7
  • On SDD, GEAR increases parameters from 3.15M to 3.18M and FLOPs from 31.44G to 31.48G relative to Resonance.The paper characterizes the added computational overhead as negligible.Source: C. Ablation Study · p. 7, Table IV · p. 7

D. Analysis of Generation-Aware Activation

GEAR varies its decoded bias activations across future steps and scene densities, and its reported gains extend across density groups and prediction horizons. The analyses relate this calibration to early motion trends and stronger interaction constraints.

D. Analysis of Generation-Aware Activation

  • In Fig. 4 cases from ETH-UCY, SDD, and NBA, GEAR predictions more closely follow ground truth than Resonance predictions.The figure compares observed history, ground truth, Resonance, and GEAR predictions.Source: Figure 4 · p. 8
  • Both bias activations are relatively low early and increase at longer horizons, as near-future motion is largely captured by the linear motion base.The analysis links later increases to the need for nonlinear self-motion and interaction-aware corrections.Source: D. Analysis of Generation-Aware… · p. 8
  • Sparse scenes show GEAR’s largest relative improvement; its lower social-bias activation there leaves early predictions closer to the linear base before corrections are introduced progressively.The paper notes that static full-strength social-bias injection may interfere with the motion trend when interaction evidence is weak.Source: D. Analysis of Generation-Aware… · p. 8
  • Dense scenes have higher self- and social-bias activations, and positive step-wise gains show calibration remains beneficial under strong interactions.The analysis interprets the stronger self-bias activation as coordination between the two components.Source: D. Analysis of Generation-Aware… · p. 8
  • GEAR consistently improves the baseline across short, middle, and full prediction horizons.The reported short-horizon ADE improvement is pronounced, and gains remain evident over the full horizon.Source: D. Analysis of Generation-Aware… · pp. 8–9, Table VI · p. 9

E. Qualitative Results

Across ETH-UCY, SDD, and NBA, qualitative examples show GEAR predictions better matching ground truth while retaining plausible motion continuity. The conclusion presents step-wise activation as a flexible, interpretable way to balance individual motion and social constraints.

E. Qualitative Results

  • Across ETH-UCY, SDD, and NBA, GEAR predictions better match ground truth while preserving plausible motion continuity.The comparison includes crowd scenarios and NBA player interactions, where GEAR adapts bias contributions across future steps.Source: E. Qualitative Results · p. 9

V. CONCLUSION

  • GEAR addresses the gap between dynamic social encoding and activation by preserving a linear base and separate bias components, then gating the decoded biases step by step.The gates modulate self-motion and social-resonance contributions before final trajectory composition.Source: §V · p. 9
  • Ablations, density analyses, activation visualizations, and qualitative results support generation-aware activation as a flexible and interpretable way to balance motion continuity with social constraints.The conclusion argues future models should consider when to strengthen or suppress social context during generation, not only how to encode it.Source: §V · p. 9

Related papers

Loading 2609.13778v1…