Abstract
Most modern multiple object tracking (MOT) systems follow the tracking-by-detection paradigm, consisting of a detector followed by a method for associating detections into tracks. There is a long history in tracking of combining motion and appearance features to provide robustness to occlusions and other challenges, but typically this comes with the trade-off of a more complex and slower implementation. Recent successes on popular 2D tracking benchmarks indicate that top-scores can be achieved using a state-of-the-art detector and relatively simple associations relying on single-frame spatial offsets -- notably outperforming contemporary methods that leverage learned appearance features to help re-identify lost tracks. In this paper, we propose an efficient joint detection and tracking model named DEFT, or "Detection Embeddings for Tracking." Our approach relies on an appearance-based object matching network jointly-learned with an underlying object detection network. An LSTM is also added to capture motion constraints. DEFT has comparable accuracy and speed to the top methods on 2D online tracking leaderboards while having significant advantages in robustness when applied to more challenging tracking data. DEFT raises the bar on the nuScenes monocular 3D tracking challenge, more than doubling the performance of the previous top method. Code is publicly available.
Introduction
Multiple Object Tracking (MOT) is a critical component for autonomous systems, typically implemented via the "tracking-by-detection" paradigm. While recent trends have favored minimalist trackers using simple spatial offsets, these methods often struggle in scenarios with low frame rates or significant occlusions. DEFT (Detection Embeddings for Tracking) addresses these limitations by integrating appearance-based re-identification directly into the detection pipeline, creating a robust and efficient joint learning framework.
The DEFT Architecture
The core innovation of DEFT lies in its ability to simultaneously perform object detection and extract descriptive appearance embeddings.
Joint Detection and Embedding Extraction
Unlike traditional two-stage trackers that use a separate network for Re-identification (ReID), DEFT leverages the intermediate multi-scale feature maps of the detector's backbone. By extracting embeddings from multiple layers, the model captures a rich hierarchical representation of objects, which is essential for accurate matching across frames.
Matching and Motion Constraints
The extracted embeddings are processed by a matching sub-network to generate an affinity matrix. To ensure physical consistency, DEFT incorporates a Long Short-Term Memory (LSTM) network that forecasts object trajectories. This motion model penalizes potential matches that deviate significantly from predicted paths, effectively combining appearance and motion cues.
Experimental Results
DEFT was evaluated on several benchmarks, including MOT16/17, KITTI, and the particularly challenging nuScenes 3D tracking dataset.
Robustness to Challenging Conditions
A key strength of DEFT is its performance in the "non-local" regime, where objects move significantly between frames due to low sampling rates or high velocities. On the nuScenes benchmark, which features a low 2 FPS annotation rate, DEFT achieved an AMOTA of 17.7, more than doubling the performance of the previous state-of-the-art, CenterTrack (4.6 AMOTA).
Efficiency and 2D Performance
On standard 2D benchmarks like MOT17 and KITTI, DEFT remains competitive with top-tier methods while significantly reducing identity switches. The joint architecture allows the system to run at approximately 12.5 Hz, making it suitable for real-time robotics applications.
Conclusion
DEFT demonstrates that joint learning of detection and appearance embeddings, augmented by motion forecasting, provides a powerful solution for robust object tracking. By moving beyond simple spatial offsets, DEFT handles long-term occlusions and low-frame-rate data far more effectively than previous methods. Future work includes extending this framework to LiDAR-based detection and incorporating ego-motion compensation to further refine trajectory predictions.
Summary updated