Citation

OmniSmall was introduced with OmniSORT:
Re-Engineering Sort-Based Algorithms for Low Cost Small Object Tracking from Omnidirectional Footage
GitHub: Xin-Shu/OmniSORT
Please cite that work when using this dataset.

Bibtex:

@misc{shu2026reengineeringsortbasedalgorithmslowcost,
      title={Re-engineering SORT-based algorithms for low-cost small object tracking from omnidirectional footage}, 
      author={Xin Shu and Meegan Gower and Yvonne Buckley and Anil Kokaram},
      year={2026},
      eprint={2609.07547},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.07547}, 
}

OmniSmall

A benchmark for small-object detection and tracking in low-cost omnidirectional footage.

14 annotated 360-degree sequences of birds and pollinating insects, captured largely on consumer 360 cameras. Objects occupy a median of 0.00055% of the frame -- roughly one pixel in 180,000 -- which puts OmniSmall well outside the size regime that mainstream detection benchmarks cover.

Why this dataset exists

Ecological monitoring has a measurement problem. Counting birds around a wind turbine, tracking insects across a flower bed, or surveying a roost currently means either a human observer, who does not scale and whose presence disturbs the subject, or a rig of pointed cameras, which is expensive and misses whatever falls outside its field of view.

A single consumer 360 camera removes both constraints: it costs a few hundred euro, records the entire hemisphere at once, and can be left unattended. What it does not remove is the analysis problem. At 8K equirectangular, a bird twenty metres away is about twelve pixels across, it may cross the projection seam mid-flight, and a single frame is 29.5 megapixels of mostly empty sky. Detectors trained on COCO-scale objects do not transfer, and tiled inference over a 30 MP frame is expensive enough to rule out continuous or on-device deployment.

OmniSmall exists so that these methods can be measured rather than asserted. Concretely it supports:

• biodiversity and avian monitoring -- unattended, non-intrusive census and behaviour recording; human annotation is difficult;
• collision-risk assessment around wind farms and airfields, where the quantity of interest is small-object flight paths near a fixed structure;
• efficient inference research -- the extreme object-to-frame ratio makes this a natural testbed for methods that decide where to spend computation.

Being consumer-camera footage matters for the first three: the barrier to conservation groups using computer vision is usually equipment cost and expertise, not algorithms.

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.

Statistics

sequences	14
annotated frames	4,377
annotated instances	41,277
unique tracks	310
mean objects per frame	9.4 (range 1.3 - 145)
resolutions	7680x3840 (10 seq), 5760x2880 (3), 5120x2560 (1)
cameras	QooCam 3 (10), Ricoh Theta X (3), broadcast footage (1)
subjects	birds; honey bees in 2 sequences

Object scale -- the defining property

percentile	p25	median	p75	p95	max
$$\sqrt{h\times w}$$	7.5	12.7	21.0	39.8	374

threshold	share of instances
< 8 px	27.6%
< 12 px	47.0%
< 16 px	62.6%
< 32 px (COCO "small")	90.4%

Per-sequence

sequence	camera	resolution	frames	tracks	instances	obj/frame	median px
bbc_earth	broadcast	5120x2560	100	171	14,511	145.1	15.5
football_pitch	QooCam 3	7680x3840	201	34	6,119	30.4	21.8
calicut_1	QooCam 3	7680x3840	500	19	5,523	11.0	7.3
dunlaoghaire_dock_1	QooCam 3	7680x3840	500	30	3,111	6.2	7.5
cricket_pitch	QooCam 3	7680x3840	200	21	2,970	14.8	7.5
clontarf	QooCam 3	7680x3840	500	7	2,783	5.6	14.1
thetax_avoca230723	Ricoh Theta X	5760x2880	500	4	1,702	3.4	5.0
qoocam_avoca230723	QooCam 3	7680x3840	500	3	1,363	2.7	9.0
Q360_20250912_121426	QooCam 3	7680x3840	200	6	823	4.1	23.1
R0010116	Ricoh Theta X	5760x2880	200	3	600	3.0	56.2
qoocam_field	QooCam 3	7680x3840	425	3	557	1.3	7.4
Q360_20250912_120202_full	QooCam 3	7680x3840	200	3	433	2.2	23.2
R0010117	Ricoh Theta X	5760x2880	200	2	394	2.0	37.2
qoocam_patio	QooCam 3	7680x3840	151	4	388	2.6	10.5

Layout

data/<sequence>/
    frame/img%04d.png     equirectangular frames
    gt.txt                MOT format: frame,id,x,y,w,h,...
    det.txt               detector output, where available

Sample 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

Licence

CC-BY-4.0.

---