FeatureSORT: Robust Multi-Object Tracking via Feature-Enriched Detection
Tracking multiple people across video frames is difficult when targets look similar, get occluded, or cross paths. Most trackers rely on bounding boxes and generic appearance features, which often fail in crowded scenes.
FeatureSORT addresses this by redesigning the detector to predict not only bounding boxes but also clothing color, clothing style, and movement direction for each person, mimicking how humans naturally track individuals in a crowd. These cues are combined with re-identification features into a unified distance function for matching detections to tracks. The result is more stable identities through occlusions and fewer identity switches. Experiments across MOT16, MOT17, MOT20, and DanceTrack show improved state-of-the-art online tracking performance without relying on offline post-processing.
Structured Temporal Inference in Hybrid State-Space Models
The second paper introduces π-SSM, a framework for inferring latent states in systems where dynamics switch between discrete regimes, such as a ball bouncing off walls or a robot changing navigation modes. Unlike classical approaches like Switching Linear Dynamical Systems (SLDS), which require hand-designed mode transition priors, π-SSM uses a neural network to infer discrete mode variables directly from the continuous latent state. Continuous states are then estimated via Kalman-style updates parameterized by a compact RNN.
A key technical contribution is a stability-aware training scheme based on input-to-state stability (ISS) theory that prevents gradient instabilities that commonly arise in recurrent architectures. The framework achieves state-of-the-art results on synthetic benchmarks (bouncing ball, Lorenz attractor, Navier-Stokes), as well as real-world robot localization, in state estimation, regime detection, and imputation under noise and partial observability.
References
[1] Hamidreza Hashempoor, Rosemary Koikara, and Yu Dong Hwang. 2026. FeatureSORT: A Robust Tracker with Optimized Feature Integration. Pattern Recognition, 175, 113148. https://www.sciencedirect.com/science/article/pii/S0031320326001111
[2] Hamidreza Hashempoor, et. al. 2026. Structured Temporal Inference in Hybrid State-Space Models. AISTATS 2026. https://drive.google.com/file/d/136DAwF_H_uuTQM3G-U2Xty_a9Qh6fx4Z/view?usp=drive_link