trajectory convolution
**Trajectory convolution** is the **motion-aligned convolution strategy that samples features along estimated object paths instead of fixed straight temporal tubes** - this improves temporal aggregation when objects move significantly across frames.
**What Is Trajectory Convolution?**
- **Definition**: Convolution operation where temporal sampling offsets follow motion trajectories.
- **Core Idea**: Align receptive field with moving content to reduce motion blur in feature space.
- **Difference from 3D Conv**: Standard 3D kernels sample fixed positions through time.
- **Input Requirement**: Motion cues from optical flow or learned offset predictors.
**Why Trajectory Convolution Matters**
- **Motion Robustness**: Better feature continuity for fast-moving objects.
- **Signal Quality**: Reduces mixing of unrelated background pixels across frames.
- **Efficiency**: Focuses computation on relevant trajectories instead of dense temporal neighborhoods.
- **Detection Gains**: Improves recognition under camera and object motion.
- **Compatibility**: Can augment existing 2D or 3D convolution backbones.
**Trajectory Modeling Approaches**
**Flow-Guided Sampling**:
- Use optical flow vectors to shift sampling coordinates over time.
- Explicitly follows estimated displacement field.
**Learned Deformable Offsets**:
- Predict offsets end-to-end for task-specific alignment.
- Avoids explicit flow supervision.
**Hybrid Schemes**:
- Start with flow prior then refine with learnable offsets.
- Balances physical consistency and task optimization.
**How It Works**
**Step 1**:
- Estimate temporal motion offsets for each spatial location across adjacent frames.
**Step 2**:
- Apply convolution using offset sampling paths that track moving structures and aggregate aligned features.
Trajectory convolution is **a motion-aware filtering method that keeps kernels locked on moving targets over time** - it is especially useful when fixed temporal sampling causes heavy misalignment artifacts.