TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations

AI Digest

TrackEverything通过3D场景表示打破点跟踪的长时稀疏与短时密集权衡，实现长时段密集跟踪并在数据集上超越现有方法。

TrackEverything breaks the tradeoff between sparse long-term and dense short-term tracking via 3D scene representations, achieving superior performance on benchmark datasets.

Key points

• 提出去重机制合并共存轨迹，避免冗余观测 Introduces deduplication mechanism to merge co-located tracks
• 分解跟踪任务为端点细化与轨迹细化两阶段 Decomposes tracking into endpoint refinement and trajectory refinement
• 3D WAFT替代4D卷积体积，提升内存效率 3D WAFT replaces memory-heavy 4D correlation volumes

Takeaway: 3D场景表示使密集跟踪突破内存限制，实现长时高效追踪。 / 3D scene representation enables dense tracking beyond memory constraints for long sequences.

Why it matters 该方法在资源受限场景下实现长时密集跟踪，对自动驾驶和机器人领域有实际应用价值。