Tracking Intermittent Particles with Self-Learned Visual Features

Jul 16, 2026

Upload by Maeve

Time-lapse fluorescence imaging relies on single-particle tracking to quantify biological dynamics, yet occlusions and intermittent detectability break continuity. Persistent gaps over a few frames often lead to multiple tracklets for one particle. This work introduces self-supervised learning to extract visual features for comparing tracklets, combining visual and positional distances to robustly stitch them. Evaluation on time-lapse fluorescence data of Hydra Vulgaris neurons shows high stitching precision and halves errors versus prior methods on the same dataset.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

Abstract

Introduction

Single-particle tracking pipeline

Particle detection and localization

Linking detections across time

Intermittent particles and tracklets

Proposed approach and contributions

© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. DOI: 10.1109/ISBI53787.2023.10230664 TRACKING INTERMITTENT PARTICLES WITH SELF-LEARNED VISUAL FEATURES Raphael Reme⋆† Victor Piriou⋆ Alison Hanson‡ Rafael Yuste‡ Alasdair Newson† Elsa Angelini† Jean-Christophe Olivo-Marin⋆ Thibault Lagache⋆ ⋆ Institut Pasteur, Universit de Paris-Cit, CNRS UMR 3691, BioImage Analysis Unit F-75015 Paris, France † LTCI, Telecom Paris, Institut Polytechnique de Paris, France ‡ Department of Biological Sciences, Columbia University, New-York, U.S.A. Corresponding author: [raphael.reme@pasteur.fr](raphael.reme@pasteur.fr) arXiv :2607 .09829v1 [ ee ss .IV] 10 Jul 2026 ABSTRACT In time-lapse fluorescence imaging, single-particle-tracking is a powerful tool to monitor the dynamics of objects of interest, and extract information about biological processes. However, tracked particles can be subject to occlusion and intermittent detectability. When these phenomena persist over a few frames, tracking algorithms tend to produce multiple tracklets for the same particle. In this work, we introduce self-supervised learning of visual features to compare tracked particles, and we exploit both visual and positional distances to robustly stitch tracklets representing the same particle. We demonstrate the performance of our stitching framework on time-lapse fluorescence sequences of Hydra Vulgaris neurons. Results show high stitching precision, and reduction of errors made by previous algorithms on the same data by a factor of two. Index Terms— Single Particle Tracking, Optimization, Deep Learning, Self-supervised Learning 1. INTRODUCTION To study the dynamics of particles (e.g., molecules, pathogens, cells...) in fluorescence imaging, single particle tracking (SPT) is required to extract meaningful information for each tracked particle. SPT is usually divided into two different stages. First, particle detection and localization is performed on each time frame. Then, the detections are linked into coherent tracks. The detection of particles’ spots in biological images can be solved in different ways. Standard approaches typically involve multiple steps including noise reduction (e.g. Gaussian smoothing, wavelet denoising...), signal enhancement ([e.g. top-hats](e.g. top-hats), h-domes...) and thresholding [1] . Newer algorithms using deep learning such as StarDist [2] or simple object detection approaches such as fasterRCNN or Yolo [3, 4] have shown better performance, but require manual annotation of spots for training. Linking detections through time is complex when numerous interacting particles are involved. In addition, the detection step usually leads to missed and false detections due to poor signal-to-noise ratio (SNR) . Therefore, elaborate tracking algorithms have been developed over the years to robustly link detections into coherent tracks for each particle. A first class of algorithms relies on global distance minimization (GDM) between consecutive frames detections [5, 6] . But frame-to-frame linking is not sufficient, and solving the GDM problem over all frames is infeasible due to memory and time limitations. Therefore, heuristic methods must be used. In [6] the frame-to-frame GDM is solved to produce tracks and GDM is reapplied globally to merge and split tracks and resolve inherent issues of frame-to-frame linking. Tracking can also be solved with proba- bilistic frameworks. For instance, using Kalman filters or Interacting Multiple Models (IMM) with likelihood maximization [7] . Multiple Hypothesis Tracking (MHT) heuristics can also be used. This keeps several hypotheses at each frame,

FAQs

Why does intermittent detectability cause multiple tracklets instead of a single track per particle?

When a particle becomes undetectable for several frames, linking methods struggle to maintain continuity and instead create multiple partial tracks (tracklets) for the same particle.

What is the key idea of the proposed tracklet stitching method?

The method builds a tracklet-to-tracklet cost using both learned visual distances and positional distances, then applies global distance minimization to optimally stitch tracklets into complete trajectories.

How are visual features for tracklets obtained?

Square image patches around detected spots are extracted and embedded into discriminative features using a self-supervised trained convolutional neural network.

Jul 16, 2026

Upload by Maeve