PlantTracing: Tracing Arabidopsis Thaliana Apex with CenterTrack

TL;DR

The paper addresses automated tracking of the Arabidopsis thaliana apex in time-lapse sequences to quantify motion and growth. It adapts CenterTrack, a CenterNet-based encoder–decoder multi-object tracker, incorporating a deformable convolutional backbone and Deep Layer Aggregation, trained on ten labeled videos and tested on three; inputs include frames $I^{(t-1)}$, $I^{(t)}$ and tracked objects $T^{t-1}$, with outputs including apex centers $p_i$, sizes $s_i$, and a displacement map $\hat{D}^t$, derived from a heat map $\hat{Y}$ down-sampled by $R=4$. Key findings show improved localization accuracy (lower MSE) compared to prior methods, with performance modulated by apex visibility and the tracking threshold, where higher thresholds reduce false positives but may miss detections. Significance lies in providing a scalable, automated workflow for high-throughput plant phenotyping and apex growth analysis.

Abstract

This work applies an encoder-decoder-based machine learning network to detect and track the motion and growth of the flowering stem apex of Arabidopsis Thaliana. Based on the CenterTrack, a machine learning back-end network, we trained a model based on ten time-lapsed labeled videos and tested against three videos.

Figures (13)

• Figure 1: Architect of Deep Layer Aggregation from deformableCNN
• Figure 2: Architect of Deep Layer AggregationDLA
• Figure 3: CenterTrack input and outputCenterTrack
• Figure 4: Training video setup
• Figure 5: Options of Backbone used in CenterNetCenterNet