xinsxins/OmniSmall

OmniSORT

Re-Engineering Sort-Based Algorithms for Low Cost Small Object Tracking from Omnidirectional Footage license: cc-by-4.0

Overview

Most multi-object trackers are designed for standard cameras and often rely on heavy appearance models. These assumptions do not work well for low-cost omnidirectional footage, where seam discontinuities and tiny fast-moving targets make tracking difficult.

This repository presents two modifications of SORT-based trackers (SORT and OCSORT) to adapt multi small objects tracking from omnidirectional footages.

Contributions

• A seam-aware Kalman filter that preserves spherical continuity across projection borders;
• OmniEuc + GIoU ($E_{fuse}$) for more robust object association in omnidirectional tracking;
• OmniSmall, a benchmark for small-object tracking in omnidirectional footage, with real-world trajectories, strong distortion, and frequent seam crossings.

---

Dataset Snapshot

Representative frames from four omnidirectional sequences illustrating object trajectories. Coloured bounding boxes accumulated over time visualise the per-object tracks in the equirectangular projection. The upper-right frame shows honey bees around a flower bed (after a 90° vertical rotation of projection), while the remaining frames show avian species. Insets show zoomed-in crops of several tracked identities, highlighting strong appearance ambiguity.

Links to dataset:
OmniSmall:

https://huggingface.co/datasets/xinsxins/OmniSmall

JRDB:

https://jrdb.erc.monash.edu/dataset/panotrack

Method

Our approach revisits SORT-style tracking for omnidirectional small-object scenarios and introduces modifications designed for limit compute resource, weak-appearance, motion-dominated tracking.

1. $Kalman\ Speed\ Update$: Seam-Aware Motion Model

Standard motion updates can break near the association at borders of an equirectangular frame, where an object may appear to jump across the seam. Our seam-aware Kalman speed update accounts for this wrap-around behaviour, giving more stable motion prediction in omnidirectional footage.

2. $OmniEuc$: Seam-Aware Euclidean Distance

Standard Euclidean distance does not reflect true proximity when objects are close across the image seam. OmniEuc fixes this by measuring distance in a seam-aware way, making object association more reliable in omnidirectional views.

3. $E_{fuse}$: Composite Association Metrics

$E_{fuse}$ combines OmniEuc with GIoU to build a stronger association cost. This helps the tracker use both seam-aware position cues and box-overlap cues when matching detections across frames.

---

Results

Demonstration video of tracking performance.
Footage: BBC Earth.
Top row: Ground Truth label.
Middle row: predicted label by 2 baselines (SORT and OCSORT).
Bottom row: predicted label by OmniSORT and OmniOCSORT.
Note: in case the video cannot rander, the source file locates at assets/video/demo_bbc_earth.mp4

Tracking Results on OmniSmall dataset

with ground-truth labels

Method	HOTA $\uparrow$	MOTA $\uparrow$	IDF1 $\uparrow$	IDSw $\downarrow$	FPS $\uparrow$
SORT	64.16	82.39	76.19	394	0.63
ByteTrack	67.95	84.71	84.74	93	1.51
OCSORT	75.30	77.51	74.10	170	1.26
HybridSORT	60.49	66.83	54.23	1154	0.61
OmniSORT + $E_{fuse}$ (ours)	90.88	97.94	94.04	18	1.17
OmniOCSORT + $E_{fuse}$ (ours)	92.21	96.20	91.12	37	0.97

With YOLOX detection labels

Method	HOTA $\uparrow$	MOTA $\uparrow$	IDF1 $\uparrow$	IDSw $\downarrow$	FPS $\uparrow$
SORT	36.43	1.91	44.85	353	0.51
ByteTrack	31.85	20.25	34.94	37	2.02
OCSORT	41.63	32.29	50.94	111	1.01
HybridSORT	40.78	32.70	49.03	183	0.66
OmniSORT + $E_{fuse}$ (ours)	35.57	2.81	43.36	394	1.08
OmniOCSORT + $E_{fuse}$ (ours)	44.19	36.44	55.61	172	1.00

---

Ablation Study

These ablation experiments evaluate the impact of each proposed component in the tracking pipeline. For OmniSORT, the baseline is SORT; for OmniOCSORT, the baseline is OCSORT. The reported p-values indicate whether the change from the corresponding baseline is statistically significant.

Takeaway. The results show that the proposed seam-aware association design improves tracking performance, especially on the omnidirectional benchmark. In general, OmniEuc and $E_{fuse}$ bring the largest gains in the target setting, while the effect of each component is more mixed on JRDB. Overall, the combined design is most effective on challenging omnidirectional small-object tracking.

Performance on OmniSmall dataset with Ground-Truth labels

Method	HOTA	MOTA	IDF1	IDSw	p-value (HOTA)	p-value (MOTA)	p-value (IDF1)
OmniSORT + IoU	19.40	18.05	14.02	2215	0.0005	0.0008	0.0006
OmniSORT + GIoU	57.81	64.14	52.90	1306	0.0683	0.0091	0.0069
OmniSORT + OmniEuc	88.88	97.82	91.02	36	0.0012	0.0285	0.0084
OmniSORT + $E_{fuse}$ ($\lambda$=0.9)	90.88	97.94	94.04	18	0.0095	0.0506	0.0037
OmniOCSORT + IoU	76.29	79.54	72.96	190	0.0725	0.2446	0.2499
OmniOCSORT + GIoU	88.98	88.92	88.13	50	0.0397	0.0485	0.0331
OmniOCSORT + OmniEuc	90.98	95.64	90.31	54	0.0049	0.0192	0.0034
OmniOCSORT + $E_{fuse}$ ($\lambda$=0.7)	92.21	96.20	91.12	37	0.0049	0.0195	0.0029

Performance on JRDB dataset with Ground-Truth labels

Method	HOTA	MOTA	IDF1	IDSw	p-value (HOTA)	p-value (MOTA)	p-value (IDF1)
OmniSORT + IoU	53.67	84.62	44.03	18346	0.0000	0.0000	0.0000
OmniSORT + GIoU	63.27	92.16	53.97	13073	0.0000	0.0093	0.0724
OmniSORT + OmniEuc	41.15	92.21	34.43	26201	0.0000	0.6314	0.0000
OmniSORT + $E_{fuse}$ ($\lambda$=0.3)	65.23	94.07	56.56	9125	0.0000	0.0000	0.0002
OmniOCSORT + IoU	67.64	90.33	60.70	5907	0.0014	0.0003	0.7380
OmniOCSORT + GIoU	63.43	91.99	57.85	7594	0.2370	0.0000	0.4088
OmniOCSORT + OmniEuc	63.22	91.93	57.75	7741	0.5108	0.0002	0.5212
OmniOCSORT + $E_{fuse}$ ($\lambda$=0.9)	64.79	91.61	58.73	6931	0.2855	0.0001	0.3082

Performance on OmniSmall dataset with YOLOX detection labels

Method	HOTA	MOTA	IDF1	IDSw	p-value (HOTA)	p-value (MOTA)	p-value (IDF1)
OmniSORT + IoU	26.97	1.47	30.31	771	0.0036	0.5235	0.0303
OmniSORT + GIoU	36.12	2.81	43.47	392	0.3461	0.1775	0.4099
OmniSORT + OmniEuc	33.32	-7.96	39.66	785	0.3045	0.2170	0.3540
OmniSORT + $E_{fuse}$ ($\lambda$=0.1)	35.57	2.81	43.36	394	0.2015	0.2322	0.2226
OmniOCSORT + IoU	42.84	34.89	52.14	105	0.0031	0.0148	0.0019
OmniOCSORT + GIoU	42.96	34.87	52.47	104	0.0038	0.0185	0.0198
OmniOCSORT + OmniEuc	43.02	36.06	53.09	194	0.0050	0.0187	0.0133
OmniOCSORT + $E_{fuse}$ ($\lambda$=0.6)	44.19	36.44	55.61	172	0.0080	0.0100	0.0042

Performance on JRDB dataset with YOLOX detection labels

Method	HOTA	MOTA	IDF1	IDSw	p-value (HOTA)	p-value (MOTA)	p-value (IDF1)
OmniSORT + IoU	22.13	16.21	20.30	38552	0.0000	0.2426	0.0000
OmniSORT + GIoU	25.26	15.89	23.83	25142	0.4844	0.1247	0.0413
OmniSORT + OmniEuc	11.70	-3.74	10.81	116906	0.0000	0.0000	0.0000
OmniSORT + $E_{fuse}$ ($\lambda$=0.1)	25.55	15.64	24.16	24030	0.8478	0.1581	0.1046
OmniOCSORT + IoU	27.03	38.22	27.79	13475	0.0000	0.0764	0.0000
OmniOCSORT + GIoU	24.15	37.50	25.48	15426	0.0000	0.0621	0.0003
OmniOCSORT + OmniEuc	22.44	36.77	23.81	17265	0.0000	0.0425	0.0000
OmniOCSORT + $E_{fuse}$ ($\lambda$=0.5)	25.55	37.95	26.96	14401	0.0000	0.0321	0.0000

---

Installation & Sample usage(TODO)

git clone https://github.com/Xin-Shu/OmniSORT.git
cd OmniSORT