---
title: 'CylindTrack: Panoramic MOT Framework'
url: https://www.emergentmind.com/topics/cylindtrack
type: topic
---

# CylindTrack: Panoramic MOT Framework

CylindTrack is a depth-aware cylindrical tracking-by-detection framework for panoramic multi-object tracking in equirectangular panoramic videos. It is formulated for settings in which the horizontal image domain is periodic rather than Euclidean, so a target crossing the \(0^\circ/360^\circ\) seam undergoes an artificial horizontal jump in UV, which breaks planar motion assumptions and makes IoU-based association unreliable near the seam. The framework addresses this together with unstable monocular depth cues by combining Depth-Temporal Trajectory Modeling (DTM), Spherical Spatio-Temporal Consistency Learning (SSTC), and a Topology-Aware Cylindrical Motion Model (TCMM), with the stated goal of improving identity preservation and trajectory continuity in challenging panoramic scenes [2606.30097].

## 1. Problem setting and design objectives

Panoramic MOT is not treated as a straightforward extension of perspective MOT. In equirectangular panoramic videos, the left and right boundaries are adjacent in \(3\)D space but separated in pixel coordinates. This periodic horizontal domain breaks planar-Kalman and IoU-based association assumptions, especially when a trajectory crosses the seam. At the same time, large-FoV scenes often contain more objects, stronger scale variation, and more frequent interactions, so online association becomes particularly sensitive to unstable frame-wise depth cues [2606.30097].

CylindTrack is explicitly designed around four goals stated in its formulation. It seeks to respect cyclic topology by lifting horizontal motion into an unwrapped angular state \(\theta \in \mathbb{R}\) \((\bmod\ 2\pi)\), to stabilize depth cues by promoting per-frame depth observations \(z_t^{\rm obs}\) into a temporally filtered trajectory-level state \(d_t\), to improve panoramic geometry alignment through SSTC, and to preserve the standard Tracking-By-Detection structure of detection \(\rightarrow\) motion prediction \(\rightarrow\) data association for ease of deployment and interpretability [2606.30097].

The framework is therefore not only a motion model. It is a composite MOT system in which topology, depth filtering, and spherical representation refinement are treated as coupled sources of association robustness. A plausible implication is that CylindTrack targets the failure modes that are specific to panoramic imagery rather than merely reusing planar MOT components on a wider field of view.

## 2. Depth-Temporal Trajectory Modeling

For each active track \(i\) at time \(t\), CylindTrack maintains two decoupled Kalman states. The cylindrical motion state is
\[
m_t^i = [\theta_t, y_t, a_t, h_t, \dot{x}_t, \dot{y}_t, \dot{a}_t, \dot{h}_t]^\top,
\]
where \(\theta_t \in \mathbb{R}\) is the horizontal angle (unwrapped), \(y_t\) is the pixel row, \(a_t\) is the box aspect ratio, and \(h_t\) is the box height. The depth state is
\[
r_t^i = [d_t, \dot{d}_t]^\top,
\]
where \(d_t\) is the filtered depth and \(\dot{d}_t\) its temporal velocity [2606.30097].

The depth filter is a constant-velocity Kalman model:
\[
\hat{r}_{t|t-1} = F \cdot r_{t-1}, \qquad
P_{t|t-1} = F \cdot P_{t-1} \cdot F^\top + Q,
\]
with
\[
F =
\begin{bmatrix}
1 & 1\\
0 & 1
\end{bmatrix}.
\]
The observation model is
\[
z_t^{\rm obs} = H \cdot r_t + \eta, \qquad H = [1\ \ 0], \qquad \eta \sim \mathcal{N}(0,R),
\]
and the update is
\[
K_t = P_{t|t-1} \cdot H^\top (H \cdot P_{t|t-1} \cdot H^\top + R)^{-1},
\]
\[
r_t = \hat{r}_{t|t-1} + K_t \cdot (z_t^{\rm obs} - H \cdot \hat{r}_{t|t-1}),
\]
\[
P_t = (I - K_t \cdot H) \cdot P_{t|t-1}.
\]

This formulation promotes instance depth from an isolated frame-wise cue to a temporally filtered trajectory-level state. The corresponding depth-guided association term between track \(i\) and detection \(j\) is
\[
C_z(i,j) = \frac{|\hat{d}_t^i - z_t^j|}{Z_{\max} + \epsilon},
\]
where \(Z_{\max} = \max |\hat{d}_p - z_q|\) over all \(p,q\) in the frame [2606.30097].

The significance of DTM lies in its treatment of depth as a dynamic quantity rather than a per-frame attribute. The paper’s own description emphasizes suppression of short-term fluctuations, particularly near occlusions or severe distortion, where raw monocular depth is unstable.

## 3. Spherical Spatio-Temporal Consistency Learning

SSTC refines depth-aware representations through two modules: a Temporal Mixer and Spherical Geometry-aware Attention. The Temporal Mixer takes depth queries \(X \in \mathbb{R}^{(B \cdot T) \times Q \times C}\), reshapes them to \(Y \in \mathbb{R}^{B \times T \times Q \times C}\), and applies a depth-query-wise \(1\)D depth mixer along time:
\[
Y' = \mathrm{LN}(Y),
\]
\[
M_t(Y') = \mathrm{PWConv}(\sigma(\mathrm{DWConv}_t(Y'))),
\]
\[
X_{\rm tm} = X + \rho \cdot \mathrm{reshape}^{-1}(M_t(Y')).
\]
In this construction, \(\mathrm{DWConv}_t\) gathers local temporal neighbors, \(\mathrm{PWConv}\) mixes channels, and \(\rho\) is a small learned scaling [2606.30097].

The second SSTC component is Spherical Geometry-aware Attention. A fixed spherical geometry context
\[
S_{\rm geo} \in \mathbb{R}^{N_s \times C}, \qquad N_s = \sum_m H_m W_m,
\]
is precomputed, with each token encoding its \((\phi,\theta)\) via Fourier embeddings. With temporally refined queries \(X_{\rm tm}\), the module forms
\[
Q = \mathrm{LN}(X_{\rm tm}) W^Q, \qquad K_{\rm geo},V_{\rm geo} = \mathrm{LN}(S_{\rm geo}) W^{KV}.
\]
It then defines a geometry modulation
\[
G_{ij} = G(\Delta \phi_{ij}, \Delta \theta_{ij}) \in \mathbb{R}^C
\]
from relative spherical offsets, and computes
\[
\alpha_{ij} = \mathrm{softmax}_j \left[\frac{(Q_i \odot G_{ij}) \cdot K_{{\rm geo},j}^\top}{\sqrt{C}}\right],
\]
\[
A_i = \sum_j \alpha_{ij} V_{{\rm geo},j},
\]
\[
X' = X_{\rm tm} + A.
\]
The module can optionally be decomposed into horizontal and vertical branches for efficiency [2606.30097].

SSTC is presented as the mechanism that enhances temporal coherence and panoramic geometric alignment in depth-aware representations. This suggests that CylindTrack does not treat depth stabilization and spherical geometry as separate post hoc corrections; instead, it integrates them at the representation level before association.

## 4. Topology-aware cylindrical motion and periodic association

TCMM is the component that makes CylindTrack specifically cylindrical. It converts the horizontal box center \(x_c \in [0,W]\) to an unwrapped angle
\[
\theta = 2\pi(x_c/W - 1/2) \in [-\pi,\pi]
\]
and then lifts that state to \(\mathbb{R}\). Seam-consistent prediction is performed by
\[
\hat{\theta}_{t|t-1} = \theta_{t-1} + \dot{\theta}_{t-1} \cdot \Delta t.
\]
When a new detection angle \(\theta_{\rm obs}\) is observed, it is lifted to the nearest equivalent around the prediction:
\[
\Delta \theta = ((\theta_{\rm obs} - \hat{\theta} + \pi) \bmod 2\pi) - \pi,
\]
\[
\theta^* = \hat{\theta} + \Delta \theta.
\]
The Kalman update is then performed on \(\theta^*\) in \(\mathbb{R}\) [2606.30097].

For association, CylindTrack defines a periodic IoU. A box is represented as \((\theta, y, \Delta \theta, h)\), with horizontal interval
\[
I_\theta = [\theta - \Delta \theta/2,\ \theta + \Delta \theta/2]_{2\pi}
\]
and vertical interval
\[
I_y = [y - h/2,\ y + h/2].
\]
The intersection area is
\[
A_\cap = |I_\theta \cap I_\theta|_{2\pi} \cdot |I_y \cap I_y|,
\]
the box area is \(A = \Delta \theta \cdot h\), and the horizontal-periodic, pixel-vertical IoU is
\[
\mathrm{IoU}_{\rm hp}(i,j) = \frac{A_\cap}{A_i + A_j - A_\cap + \epsilon}.
\]
The overlap cost is
\[
C_{\rm hp}(i,j) = 1 - \mathrm{IoU}_{\rm hp}(i,j).
\]

The joint association cost is then constructed in stages. The spatial-depth cost is
\[
C_{d\text{--}hp}(i,j) = (1-\lambda_z) C_{\rm hp}(i,j) + \lambda_z C_z(i,j),
\]
the angular consistency cost is
\[
C_\theta(i,j) = \frac{|\,\mathrm{wrap}(\theta_i - \theta_j)\,|}{\pi},
\qquad
\mathrm{wrap}(\delta) = ((\delta + \pi) \bmod 2\pi) - \pi,
\]
and the combined cost is
\[
C_{\rm assoc}(i,j) = (1-\lambda_\theta) C_{d\text{--}hp}(i,j) + \lambda_\theta C_\theta(i,j).
\]
With detection confidence \(s_j\), the final form is
\[
\hat{C}(i,j) = 1 - (1 - C_{\rm assoc}(i,j)) \cdot s_j.
\]
Matching is solved by Hungarian optimization subject to one-to-one constraints [2606.30097].

The tracking loop uses two-stage matching and explicit track management. High-confidence detections with \(s_j \ge \tau_h\) are used in the first-stage match; unmatched tracks are then matched against low-confidence detections with \(\tau_l \le s_j < \tau_h\) using only \(C_{\rm hp}\) as a rescue stage. New tracks are initialized from unmatched high-confidence detections if \(s_j \ge \tau_{\rm init}\); tracks are deleted after \(B_{\rm lost}\) consecutive unmatched frames; and near duplicates are suppressed by retaining the longer-lived [2606.30097].

## 5. Empirical performance

CylindTrack is evaluated on QuadTrack and JRDB using HOTA, AssA, IDF1, MOTA, and FPS, with BCIC defined as Boundary Crossing Identity Consistency and described as measuring identity stability in a window around seam crossings [2606.30097].

| Benchmark | Reported result |
|---|---|
| QuadTrack | Full CylindTrack\(_{\rm TBD}\): HOTA 33.67 \((+14.11\ \text{vs OmniTrack})\), IDF1 40.45, AssA 34.67, BCIC 54.96, 28.6 FPS |
| JRDB | CylindTrack\(_{\rm TBD}\): HOTA 31.12, IDF1 34.33, AssA 31.35, BCIC 35.37, 21.3 FPS; gains \(+4.36\) HOTA and \(+4.62\) IDF1 over OmniTrack\(_{\rm TBD}\) |

The ablation study attributes distinct gains to each component. DTM alone improves IDF1 by \(\sim 2.9\) points, from \(34.1 \rightarrow 36.5\). SSTC, defined here as Temporal Mixer \(+\) SGA, yields an additional \(\sim 2.9\) IDF1 gain. TCMM, consisting of cylindrical Kalman \(+\) periodic IoU \(+\) angle cost, adds \(\sim 3.0\) IDF1. The full combination achieves \(+6.3\) HOTA and \(+8.2\) IDF1 over the best depth-aware detector alone [2606.30097].

The qualitative observations are aligned with the intended failure modes. At seam crossing, tracks remain continuous across \(0^\circ/360^\circ\) without fragmentation. Under heavy occlusion, temporally filtered depth separates overlapping targets and reduces ID switches. The source code is stated to be released at `https://github.com/warriordby/CylindTrack` [2606.30097].

## 6. Nomenclature, scope, and related uses of the term

The name “CylindTrack” is used in more than one research context, and the panoramic MOT framework should not be conflated with those other usages. In medical image analysis, “Graph-Based Small Bowel Path Tracking with Cylindrical Constraints” models the small bowel piecewise as a chain of finite cylinders and reports improvements of \(6.6\%\) and \(17.0\%\), in terms of tracked length, for two settings related to small bowel segmentation [2207.14436]. In particle-tracking reconstruction for CGEM-IT and ODC, the term is also used in a description of cluster-pattern selection and helix-based matching, where track-segment-finding efficiency is reported as \(99.1\%\) to \(99.7\%\) and overall CGEM–ODC matching efficiency as \(\sim 99\%\) [1604.03102].

The label is also adjacent, but not identical in meaning, to the ALICE program on a “truly cylindrical inner tracker.” Those papers concern wafer-scale Monolithic Active Pixel Sensors in a \(65\,\mathrm{nm}\) CMOS Image Sensor process, thinned and bent around the beam pipe, with material-budget reductions such as \(0.36\%\,X_0 \rightarrow 0.05\%\,X_0\) or \(0.36\%\,X_0 \rightarrow 0.09\%\,X_0\) per layer depending on the design stage, rather than panoramic multi-object tracking [2212.03165] [2307.08632] [2512.08748].

Accordingly, within current arXiv usage, CylindTrack most precisely denotes the panoramic MOT framework introduced in 2026, whereas earlier or parallel usages invoke cylindrical constraints, cylindrical detectors, or attractive cylinders in unrelated tracking problems. A plausible implication is that the term functions more as a geometric motif than as a single lineage across fields.

Source: https://www.emergentmind.com/topics/cylindtrack