DeepSORT<T>

Models & Types · AiDotNet.ComputerVision.Tracking

DeepSORT (Deep SORT) tracking with appearance features.

Inheritance: ObjectTrackerBase<T> → DeepSORT<T>

For Beginners

DeepSORT extends SORT by adding deep appearance features (Re-ID embeddings) to improve association accuracy, especially for occlusions and ID switches. It uses a cascade matching strategy.

How It Works

Key features:

  • Deep appearance descriptor (Re-ID network)
  • Cascade matching (prioritize recent tracks)
  • Combined motion and appearance matching
  • Gallery of appearance features per track

This follows the official deep_sort tracker: an xyah Kalman filter, a matching cascade by time since update with the nearest-neighbour cosine metric (budget 100, threshold MaxCosineDistance) gated at the chi-square 0.95 Mahalanobis quantile, then IoU matching (max IoU distance 0.7) for unconfirmed tracks and tracks missed for one frame, all with optimal assignment. Without appearance features the cascade uses the gated IoU distance instead of the cosine metric.

Reference: Wojke et al., “Simple Online and Realtime Tracking with a Deep Association Metric”, ICIP 2017

Constructors

Constructor Summary
DeepSORT(TrackingOptions<T> options) Creates a motion-only DeepSORT tracker: the matching cascade uses the gated IoU distance.
DeepSORT(TrackingOptions<T> options, Func<Tensor<T>, Tensor<T>> appearanceEncoder) Creates a new DeepSORT tracker whose appearance descriptors come from appearanceEncoder, which maps a [1, C, 128, 64] crop to a feature vector (the paper uses a wide residual network trained for person re-identification).

Properties

Property Summary
string Name { get; }

Methods

Method Summary
TrackingResult<T> Update(List<Detection<T>> detections)
TrackingResult<T> Update(List<Detection<T>> detections, Tensor<T> image)
void Reset()