SMOT4SB: UAV Dataset for Bird Tracking

• SMOT4SB is a specialized UAV video dataset for small multi-object tracking, labeling all birds with bounding boxes and unique IDs.
• It comprises 211 video sequences with 108,192 frames and over 370K annotated bird instances, capturing free 3D motion of both cameras and targets.
• Benchmark challenges include motion entanglement, occlusions, and weak appearance cues, driving innovations in motion- and geometry-based tracking methods.

SMOT4SB is a UAV video dataset and benchmark for Small Multi-Object Tracking for Spotting Birds, introduced in the context of the MVA 2025 challenge as a follow-up to SOD4SB, which addressed single-frame small object detection rather than tracking. Its task formulation is single-class MOT: all instances are labeled as “bird”, each bird is localized with a bounding box, and identities are maintained across frames. The dataset is explicitly designed for a regime in which both the camera and the targets move freely in 3D, producing the “motion entanglement” that the papers identify as a central source of difficulty. Reported scale is 211 video sequences, 108,192 annotated frames, 371,690 annotated bird instances, and 2,240 unique tracking IDs (Kondo et al., 17 Jul 2025).

1. Nomenclature and lineage

The acronym SMOT4SB refers to Small Multi-Object Tracking for Spotting Birds. Within the cited literature, it is tied to the MVA 2025 Small Multi-Object Tracking for Spotting Birds Challenge, its associated dataset, and the evaluation framework built around it (Kondo et al., 17 Jul 2025). A recurrent source of confusion is the earlier acronym SOD4SB, which denotes Small Object Detection for Spotting Birds, a detection benchmark rather than a tracking dataset. The relationship is sequential: SOD4SB focused on single-frame bird detection, while SMOT4SB extends the problem to temporally consistent multi-object tracking in UAV video (Kondo et al., 2023).

This lineage matters because the motivation shifts from framewise localization to identity-preserving temporal analysis. The challenge paper states that SMOT4SB was created to leverage temporal information precisely because bird spotting from drones is intrinsically temporal: birds are often tiny, blurred, distant, and visually ambiguous in any one frame, so single-frame detection is unreliable (Kondo et al., 17 Jul 2025). A plausible implication is that SMOT4SB should be understood less as a generic MOT benchmark than as a deliberately specialized successor to a small-object detection benchmark in the same domain.

2. Dataset composition, acquisition, and annotation

The dataset is split at the video level, so no sequence appears in more than one partition. The reported split is Train: 128 videos, 66,602 frames, 226,292 instances, 1,256 IDs; Public test: 38 videos, 16,489 frames, 51,325 instances, 509 IDs; Private test: 45 videos, 25,101 frames, 94,073 instances, 475 IDs; Total: 211 videos, 108,192 frames, 371,690 instances, 2,240 IDs (Kondo et al., 17 Jul 2025). The videos were captured at 30 fps, with image resolutions of 1920 × 1080 and 3840 × 2160 (Kondo et al., 17 Jul 2025).

Collection used on-drone cameras mounted on DJI Mavic 2 Pro, DJI Phantom 4 Pro V2.0, and ProDrone PD4B-M. The recorded environments include urban areas, parks, forests, rivers, and agricultural fields, and the paper states that capture occurred under different weather and lighting conditions (Kondo et al., 17 Jul 2025). Although the semantic label space is collapsed to a single class, the observed birds are said to include hawks, crows, waterfowl, sparrows, and various small passerine species (Kondo et al., 17 Jul 2025).

Annotation was performed manually using VATIC and CVAT. Annotators enclosed each bird with a bounding box and assigned a unique tracking ID, and the annotations were double-checked for quality (Kondo et al., 17 Jul 2025). The released annotation format is COCO format, and the papers mention only three label components: bounding boxes, track IDs, and the class label “bird”. They do not report visibility flags, occlusion flags, truncation labels, ignored regions, confidence scores in ground truth, or explicit temporal event annotations beyond per-frame boxes and identity tracks (Kondo et al., 17 Jul 2025).

One reporting ambiguity should be noted. The challenge paper states 108,192 annotated frames in the abstract and split table, but also contains a later statement giving 183,192 annotated frames; because the abstract and tables agree on 108,192, that figure is the one the paper itself supports most consistently (Kondo et al., 17 Jul 2025).

3. Task regime and characteristic difficulty

SMOT4SB targets a particularly hard subset of MOT: small-object, single-class, UAV-based tracking with weak appearance cues. The papers repeatedly state that most birds occupy fewer than $32 \times 32$ pixels, placing the benchmark in the COCO small-object regime (Kondo et al., 17 Jul 2025). At that scale, appearance descriptors are weak, detector confidence is unstable, and even very small localization errors can cause large changes in overlap-based scores.

The benchmark’s defining difficulty is motion entanglement. Unlike fixed-camera pedestrian MOT, or UAV tracking settings in which targets are constrained by roads or ground planes, SMOT4SB contains moving UAV cameras and independently moving birds, with both exhibiting free 3D motion (Kondo et al., 17 Jul 2025). The papers also emphasize flocking dynamics, including coordinated motion, trajectory crossings, visual merging and separation, mutual occlusions, identity switches, motion blur, rapid scale changes, and complex backgrounds (Kondo et al., 17 Jul 2025). One challenge solution additionally notes false positives around power poles in urban areas, which is presented as an example of hard background structure rather than a dataset-wide statistic (Kondo et al., 17 Jul 2025).

A later method paper provides a more specific motion analysis. It reports that the percentage of samples whose velocity ratio stays in $[0.8, 1.2]$ relative to an average over previous frames is 76.2% for reference window $N=1$, 71.9% for $N=2$, 67.3% for $N=3$, 64.2% for $N=4$, and 61.2% for $N=5$ (Yu et al., 16 Jul 2025). This suggests that short-term motion regularity is present often enough to remain useful, even though the overall regime is explicitly described as irregular and difficult.

4. Benchmark protocol and evaluation methodology

The challenge protocol has two phases. In the public test phase, participants receive public test videos/images without annotations, submit results to CodaBench, and obtain scores on the public subset; submission frequency is limited to 10 evaluations per day per team to reduce HARKing. In the private test phase, participants submit code, and organizers run that code on the private test subset; final rankings are based on the private test results. After the challenge, CodaBench remains publicly available for evaluation on the public test subset (Kondo et al., 17 Jul 2025).

The official baseline is a tracking-by-detection pipeline using YOLOX fine-tuned on SMOT4SB for bird detection and OC-SORT for multi-object tracking (Kondo et al., 17 Jul 2025). A separate solution paper describes the same partition sizes as 128 training sequences, 38 validation sequences, and 45 testing sequences (Yu et al., 16 Jul 2025). Because those counts match the challenge’s train / public test / private test partition exactly, a plausible implication is that the later paper treats the public test split as a de facto validation set for method reporting.

A central methodological contribution associated with SMOT4SB is the replacement of IoU-based MOT similarity with Dot Distance (DotD) inside the HOTA framework. The challenge paper defines

$\text{DotD}(A,B)=e^{-\frac{D}{S}},$

where

$D=\sqrt{(x_A-x_B)^2+(y_A-y_B)^2}$

is the center distance, and $S$ is the average object size over the dataset (Kondo et al., 17 Jul 2025). The motivation is that IoU is too sensitive to small displacements when objects are tiny: for a $[0.8, 1.2]$0 object, an 8-pixel shift can reduce IoU from 1.0 to 0.5, and once boxes no longer overlap, IoU becomes zero even if the prediction is only slightly displaced (Kondo et al., 17 Jul 2025). SO-HOTA therefore substitutes DotD for IoU in the detection and association components of HOTA, while retaining one-to-one Hungarian matching and averaging over thresholds from 0.05 to 0.95 (Kondo et al., 17 Jul 2025).

5. Baselines, challenge outcomes, and what they reveal about the dataset

The challenge attracted 78 participants and 308 submissions, with 7 teams reaching final private-test evaluation. Ranking was based on SO-HOTA, while the reported accompanying metrics included SO-DetA, SO-AssA, HOTA, MOTA, IDF1, MT, and ML (Kondo et al., 17 Jul 2025). On the private test subset, the top reported results were DL Team with SO-HOTA 50.59, SO-DetA 47.27, SO-AssA 54.30, HOTA 36.74, MOTA 28.80, and IDF1 44.86; zhwa2003 with SO-HOTA 46.22; and elsalabA with SO-HOTA 43.87 (Kondo et al., 17 Jul 2025). The official baseline scored SO-HOTA 9.90, HOTA 6.51, MOTA 1.61, and IDF1 5.18, so the winning private-test score was reported as approximately 5.1× the baseline (Kondo et al., 17 Jul 2025).

These results are used in the paper to characterize SMOT4SB as unsaturated and methodologically demanding. The large gap between SO-HOTA and ordinary HOTA across top systems is presented as evidence that IoU-based HOTA undervalues performance on tiny objects (Kondo et al., 17 Jul 2025). The same section also notes recurring design patterns among successful entries: patch/slice-based training or inference, detector ensembling, false-positive filtering, motion compensation for UAV ego-motion, and appearance-free or weak-appearance tracking (Kondo et al., 17 Jul 2025).

A later paper, describing a championship-winning public-test solution, reports SO-HOTA 55.205, SO-DetA 51.716, and SO-AssA 59.082 on the SMOT4SB public test set using YOLOv8-L within a tracking-by-detection pipeline built around SliceTrain and an appearance-free OC-SORT variant (Yu et al., 16 Jul 2025). Because these figures are from the public test rather than the private final ranking, they are not directly interchangeable with the private-test leaderboard. They are, however, informative about the dataset’s operating regime: the method paper argues that SMOT4SB makes appearance-based association unusually weak, so robust performance depends on improving small-object detection and using motion- and geometry-based matching instead (Yu et al., 16 Jul 2025).

6. Limitations, reporting gaps, and interpretive issues

SMOT4SB is deliberately narrow in semantic scope. It has 1 class, and all instances are labeled simply as “bird”; there are no species labels despite the diversity of observed birds (Kondo et al., 17 Jul 2025). The papers also do not report fine-grained metadata such as altitude, weather tags, difficulty labels, camera pitch statistics, crowd-density annotations, visibility flags, truncation labels, or explicit occlusion indicators (Kondo et al., 17 Jul 2025). Likewise, they do not provide a full object-size distribution, average trajectory length, motion-magnitude distribution, or formal annotation-density policy, although the wording strongly indicates that all frames in each included sequence are annotated (Kondo et al., 17 Jul 2025).

The annotation process is described as manual and double-checked, but the challenge paper does not quantify annotation noise (Kondo et al., 17 Jul 2025). A cautious reading is warranted because annotation of tiny flying birds under blur and occlusion is inherently difficult. The paper itself notes that maintaining identity consistency was hard in the presence of occlusion, motion blur, and scale variation, and that annotators relied on magnification and careful frame-by-frame comparison (Kondo et al., 17 Jul 2025). This suggests that marginal score differences should be interpreted carefully unless they are large relative to expected annotation uncertainty.

Finally, the literature around SMOT4SB contains two small but consequential ambiguities: the already noted 108,192 versus 183,192 frame count discrepancy, and the inconsistent naming of the non-training partitions as public/private test in the challenge paper versus validation/testing in one solution paper (Kondo et al., 17 Jul 2025, Yu et al., 16 Jul 2025). Neither ambiguity changes the core identity of the dataset. Across sources, SMOT4SB is consistently represented as a benchmark for tracking small birds in UAV videos under motion-entangled 3D conditions, with dataset scale fixed at 211 sequences and evaluation centered on SO-HOTA rather than conventional IoU-based MOT metrics (Kondo et al., 17 Jul 2025).