Core Concepts
Proposing MotionTrack, a novel tracker with a learnable motion predictor solely based on object trajectory information.
Abstract
The content introduces MotionTrack, a new online tracking approach centered around a learnable motion predictor. It addresses challenges in multi-object tracking by focusing on object trajectory information. The proposed method integrates self-attention and Dynamic MLP to enhance motion modeling. Experimental results show superior performance on datasets like Dancetrack and SportsMOT.
Structure:
- Introduction to Multi-Object Tracking
- Challenges in Object Tracking with Homogeneous Appearance and Heterogeneous Motion
- Proposed Approach: MotionTrack with Learnable Motion Predictor
- Methodology: Transformer-based Motion Prediction and Dynamic MLP Integration
- Data Augmentation Strategies: Random Drop, Spatial Jitter, Random Length
- Experiments on Datasets: DanceTrack and SportsMOT
- Ablation Studies on Model Components, Interactions, Data Augmentations, and Hyperparameters
Stats
Our experimental results demonstrate that MotionTrack yields state-of-the-art performance.
The proposed method achieves a higher HOTA score compared to OC SORT.
Using Random drop and Spatial jitter augmentation techniques improves tracking performance.