[Paper Review] DEFT: Detection Embeddings for Tracking

Mohamed Chaabane, Peter Zhang+2 more|arXiv (Cornell University)|Feb 3, 2021

Anomaly Detection Techniques and ApplicationsComputer Science53 references52 citations

TL;DR

DEFT jointly learns detection embeddings within the detector backbone to perform online multi-object tracking, using an appearance-based matching head and a motion model to robustly track objects across frames, especially under occlusion and large inter-frame displacement; it achieves strong results on 2D benchmarks and substantially boosts monocular 3D tracking on nuScenes.

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.

Motivation & Objective

• Motivate tracking-by-detection with a simpler yet robust association mechanism.
• Develop a joint detection-and-tracking network that reuses detector features for embedding-based matching.
• Incorporate a motion model to constrain plausible object trajectories during association.
• Evaluate DEFT across 2D and 3D tracking benchmarks to demonstrate robustness under occlusion and large inter-frame displacement.

Proposed method

• Extract appearance embeddings from multiple detector backbone feature maps for each detected object.
• Train a shared detection and matching network so that embeddings optimize both detection and cross-frame association.
• Use a matching head to compute pairwise similarities between current detections and track embeddings via a 1x1 convolutional network.
• Maintain a memory of track embeddings over recent frames to enable long-range association and occlusion handling.
• Apply an LSTM-based motion forecasting module to constrain plausibility of associations and filter implausible matches.
• Perform online data association with the Hungarian algorithm, incorporating non-match scores to handle new or departed objects.

Experimental results

Research questions

• RQ1 Can detector-backbone embeddings be effectively reused for appearance-based data association in online MOT?
• RQ2 Does joint detection-and-tracking training improve both detection quality and tracking robustness compared to separate stages?
• RQ3 How does a learned motion model (LSTM) interact with embedding-based matching in challenging scenarios like occlusions and large inter-frame displacements?

Key findings

• DEFT’s joint training yields competitive 2D tracking performance across MOT and KITTI benchmarks.
• DEFT substantially improves robustness to occlusions and large inter-frame displacements, outperforming prior methods on challenging data (notably nuScenes).
• The learned detection embeddings provide a strong signal for cross-frame identity association while maintaining efficiency comparable to simpler trackers.
• An LSTM motion model offers additional gains, particularly in harder sequences, and can outperform Kalman-filter-like approaches in this context.
• Across benchmarks, DEFT demonstrates that sharing features between detection and matching can surpass methods that treat detection and association separately.