Selective Learning Framework Pipeline Overview
Figure 3: Overview of our Selective Learning framework. Top: The end-to-end pipeline processing augmented trajectories (via scaling, rotation, translation) through the Scene Context Encoder and SceneSelect to route inputs to the optimal model. Bottom: Detailed modules. (a) Scene Context Encoder translates raw features into a high-dimensional sparse matrix via random projection and sparsification, grouped into pseudo-labels via clustering. (b) SceneSelect trains an additive tree-based classifier (bounded by cross-entropy loss) to predict scene labels. (c) Scene-Model Pool generates a Scheduling
Paper context
Paper title: SceneSelect: Selective Learning for Trajectory Scene Classification and Expert Scheduling Abstract: Accurate trajectory prediction is fundamentally challenging due to high scene heterogeneity - the severe variance in motion velocity, spatial density, and interaction patterns across different real-world environments. However, most existing approaches typically train a single unified model, expecting a fixed-capacity architecture to generalize universally across all possible scenarios. This conventional model-centric paradigm is fundamentally flawed when confronting such extreme heterogeneity, inevitably leading to a severe generalization gap, degraded accuracy, and massive computational waste. To overcome this bottleneck, rather than refining restricted model-centric architectures, we propose selective learning, a novel scene-centric paradigm. It explicitly analyzes the characteristics of the underlying scene to dynamically route inputs to the most appropriate expert models. As a concrete implementation of this paradigm, we introduce SceneSelect. Specifically, SceneSelect utilizes unsupervised clustering on interpretable geometric and kinematic features to discover a latent scene taxonomy. A highly decoupled classification module is then trained to assign real-time inputs to these scene categories, and a highly extensible, plug-and-play scheduling policy automatically dispatches the trajectory sequence to the optimal expert predictor. Crucially, this decoupled design ensures excellent generalization capabilities, allowing seamless integration with different off-the-shelf mod Passages referencing this figure: riables [ 32 ] . Models optimized for sparse highways fail in dense crowds; models for complex reasoning waste computation where simpler heuristics suffice [ 40 ] . This manifests as reliability and resource problems: no single model performs well everywhere, yet deploying multiple models incurs prohibitive costs [ 3 ] . The optimal architecture and capacity vary fundamentally across scene types. Figure 1: Empirical analysis on ETH-UCY. (a) Accuracy vs. efficiency trade-off across varying degrees of scene heterogeneity: lightweight models excel in sparse scenes, while Transformers dominate complex ones. (b) Target trajectories naturally partition into distinct scene clusters via PCA based on motion velocity, spatial density, and interaction patterns, revealing systematic scene heterogeneit rying degrees of scene heterogeneity: lightweight models excel in sparse scenes, while Transformers dominate complex ones. (b) Target trajectories naturally partition into distinct scene clusters via PCA based on motion velocity, spatial density, and interaction patterns, revealing systematic scene heterogeneity. Empirically, this phenomenon is confirmed across diverse datasets. As illustrated in Figure 1 (a), evaluating representative architectures across varying degrees of scene heterogeneity reveals an inherent accuracy-efficiency trade-off. Lightweight models offer minimal latency but suffer high errors in historically complex interactions, whereas Transformer architectures excel in dense environments but waste abundant computation in simpler ones. Thus, single-model deployment inevita ant computation in simpler ones. Thus, single-model deployment inevitably causes either severe performance degradation or resource waste. Meanwhile, full ensembles are computationally prohibitive fo
