SMOT4SB: UAV Small Bird Tracking Challenge

• SMOT4SB Challenge is a multi-object tracking benchmark for UAV videos of wild birds, extending single-frame detection to continuous identity maintenance.
• It introduces SO-HOTA, a custom metric that employs Dot Distance instead of IoU to robustly evaluate detection and association of small, fast-moving targets.
• The challenge promotes innovative methods that leverage motion cues and training-time slicing to enhance bird monitoring under diverse environmental conditions.

The MVA 2025 Small Multi-Object Tracking for Spotting Birds (SMOT4SB) Challenge is an international benchmark competition for small multi-object tracking in UAV videos of wild birds. It extends the MVA2023 SOD4SB challenge from single-frame small object detection to full multi-object tracking in video, and couples a dedicated dataset with a small-object-specific evaluation metric, SO-HOTA, to study detection and identity maintenance when most targets occupy only a few dozen pixels and both camera and birds move freely in 3D (Kondo et al., 17 Jul 2025, Kondo et al., 2023).

1. Origins, scope, and problem formulation

SMOT4SB was introduced as the successor to the MVA2023 SOD4SB challenge. The earlier challenge treated each frame as an independent image for single-class bird detection, whereas SMOT4SB reformulates the setting as full multi-object tracking: for every frame in a video sequence, a system must detect all birds and maintain consistent identity labels over time. The challenge was explicitly designed to exploit temporal information and motion cues in a regime where appearance is weak and conventional re-identification is unreliable (Kondo et al., 17 Jul 2025, Kondo et al., 2023).

The problem setting is unusually difficult because the targets are small birds seen from UAVs. Most birds occupy fewer than $32 \times 32$ pixels, often around $16 \times 16$, so color, texture, and fine shape provide little usable identity information. The challenge paper characterizes the scene dynamics as “motion entanglement”: both the UAV camera and the birds move freely in 3D, producing nonlinear image-plane trajectories that are not well described by fixed-camera or ground-plane assumptions. Flocking behavior adds dense local interactions, heavy mutual occlusions, crossing trajectories, and frequent identity ambiguity. Environmental factors such as forests, rivers, fields, urban scenes, haze, clouds, changing sunlight, and clutter from structures such as power poles further complicate both detection and association (Kondo et al., 17 Jul 2025).

The intended application space is practical UAV-based bird monitoring. The challenge materials highlight bird strike avoidance, agriculture and fisheries protection, and ecological or biodiversity monitoring as motivating domains. In these settings, track-level continuity matters: trajectory information supports autonomous avoidance, early warning, population surveys, migration monitoring, and behavioral analysis in ways that isolated detections do not (Kondo et al., 17 Jul 2025).

2. Dataset design and annotation protocol

The SMOT4SB dataset is the first large-scale collection specifically designed for small bird tracking in UAV scenarios with motion entanglement. It contains 211 UAV video sequences, 108,192 annotated frames, 371,690 bird instances, and 2,240 unique tracked birds. Videos are recorded at 30 fps and include both $1920 \times 1080$ and $3840 \times 2160$ resolutions. Acquisition platforms include the DJI Mavic 2 Pro, DJI Phantom 4 Pro V2.0, and ProDrone PD4B-M. The scenes cover urban areas, parks, forests, rivers, and agricultural fields under diverse daytime weather and lighting conditions. Bird species include hawks, crows, waterfowl, sparrows, and various small passerines, but all are labeled under a single class, "bird" (Kondo et al., 17 Jul 2025).

Annotations were produced with VATIC and CVAT. Each visible bird receives a bounding box and a track ID that is maintained across frames. Annotators examined sequences frame by frame, often magnifying patches, to preserve identity continuity across occlusions, motion blur, and scale change. All annotations were double-checked. Release format follows COCO, with track IDs stored in an extended field (Kondo et al., 17 Jul 2025).

The official split is made at video level, so there is no overlap between train, public test, and private test sequences.

Subset	Videos	Frames
Train	128	66,602
Public test	38	16,489
Private test	45	25,101

The train split contains 226,292 instances and 1,256 IDs; the public test split contains 51,325 instances and 509 IDs; the private test split contains 94,073 instances and 475 IDs. This construction was meant to stress four properties simultaneously: motion entanglement, flocking dynamics, scale variation with information scarcity, and visual challenges such as motion blur and background clutter (Kondo et al., 17 Jul 2025).

3. Task definition and SO-HOTA evaluation

The task input is a raw video sequence; the required output is, for every frame, a set of detections with bounding boxes, confidence scores, and integer track IDs. Public test evaluation is submission based, while final ranking is determined on a private test set by organizer-run code. The challenge imposed a limit of 10 submissions per day on the public test to reduce leaderboard overfitting (Kondo et al., 17 Jul 2025).

The core methodological contribution of the challenge is SO-HOTA, a small-object variant of HOTA. Standard HOTA balances detection accuracy and association accuracy, but it uses IoU for matching. For tiny boxes, IoU is unstable: a small localization shift can drop IoU below threshold, and non-overlapping but nearby boxes become indistinguishable from grossly incorrect predictions. SMOT4SB therefore replaces IoU with Dot Distance (DotD), a center-distance similarity measure that decays smoothly with displacement relative to a characteristic object scale (Kondo et al., 17 Jul 2025).

Let predicted and ground-truth box centers be $(x_A, y_A)$ and $(x_B, y_B)$. Then the center distance is

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

If $w_{ij}$ and $h_{ij}$ denote box width and height over the dataset, the characteristic scale is

$$S = \sqrt{\frac{\sum_{i=1}^{M}\sum_{j=1}^{N_i} w_{ij} h_{ij}}{\sum_{i=1}^{M} N_i}}.$$

Dot Distance is defined as

$16 \times 16$0

For a threshold $16 \times 16$1, a predicted box and ground truth are considered matched if $16 \times 16$2. SO-HOTA then uses the HOTA decomposition with DotD-based matching:

$16 \times 16$3

and the overall score averages over 19 thresholds:

$16 \times 16$4

This design changes the interpretation of tracking quality in an important way. A prediction that is slightly offset from a tiny bird can still receive meaningful credit if its center remains close, while predictions that drift far away are still penalized. The challenge also reports SO-DetA and SO-AssA, as well as conventional HOTA, MOTA, IDF1, MT, and ML for comparison with standard MOT practice (Kondo et al., 17 Jul 2025).

4. Baseline, participation, and official leaderboard

The organizers released a tracking-by-detection baseline composed of a YOLOX detector fine-tuned on the SMOT4SB training set and an OC-SORT tracker. OC-SORT is a motion-based tracker with Kalman filtering and IoU-based association; in the baseline it relies mainly on geometry and motion rather than appearance. On the private test set, the baseline achieved SO-HOTA $16 \times 16$5, SO-DetA $16 \times 16$6, SO-AssA $16 \times 16$7, HOTA $16 \times 16$8, MOTA $16 \times 16$9, IDF1 $1920 \times 1080$0, MT $1920 \times 1080$1, and ML $1920 \times 1080$2. On public test, it achieved SO-HOTA $1920 \times 1080$3, SO-DetA $1920 \times 1080$4, and SO-AssA $1920 \times 1080$5. The official interpretation was that the baseline had high precision but very low recall and almost no mostly tracked trajectories (Kondo et al., 17 Jul 2025).

The competition attracted 78 registered participants and 308 submissions to the public leaderboard. Seven teams submitted final runnable solutions for private test evaluation. Ranking was determined by SO-HOTA (Kondo et al., 17 Jul 2025).

Rank	Team	SO-HOTA
1	DL Team	50.59
2	zhwa2003	46.22
3	elsalabA	43.87
4	sgm	43.71
5	xmba15	40.49

The winning private-test score, 50.59 by DL Team, corresponds to a 5.1x improvement over the baseline. Additional leaderboard metrics show that the top methods improved both detection and association, rather than trading one against the other. Resource reporting also mattered operationally: DL Team used a 44M-parameter model with 0.175 s/img on an RTX 3090, while other finalists often used substantially larger models or slower pipelines (Kondo et al., 17 Jul 2025).

5. Winning method and broader methodological trends

DL Team’s method, later detailed as “YOLOv8-SMOT: An Efficient and Robust Framework for Real-Time Small Object Tracking via Slice-Assisted Training and Adaptive Association” (Yu et al., 16 Jul 2025), is a tracking-by-detection pipeline with targeted changes on both the detection and tracking sides. The detector is a YOLOv8-based small-object detector trained with a framework called SliceTrain. SliceTrain uses deterministic full-coverage slicing and slice-level stochastic augmentation to address what the authors call the “Resolution–Diversity Dilemma”: full $1920 \times 1080$6 training images preserve the tiny birds, but direct training at that resolution severely limits batch size and sample diversity. In the reported SMOT4SB configuration, images are sliced into $1920 \times 1080$7 tiles with 20% overlap, yielding a batch size increase from 1 to 6 on a single RTX 3090, while inference is still performed directly on full-resolution frames with no slicing overhead (Yu et al., 16 Jul 2025).

The tracking side is explicitly appearance-free. It builds on OC-SORT, but adds two challenge-specific components: motion direction maintenance via an exponential moving average with $1920 \times 1080$8, and an adaptive similarity metric that combines bounding box expansion by a factor of 2 with a center-distance penalty. The aim is to stabilize association when birds are too small for Re-ID and when adjacent-frame boxes may be close but non-overlapping. On the public test set, the full system achieved SO-HOTA $1920 \times 1080$9, SO-DetA $3840 \times 2160$0, and SO-AssA $3840 \times 2160$1. Ablations on the same public benchmark showed a progression from default OC-SORT at SO-HOTA $3840 \times 2160$2, to $3840 \times 2160$3 with EMA, $3840 \times 2160$4 with bounding-box expansion, and $3840 \times 2160$5 after adding the distance penalty. The same paper reported YOLOv8-L, YOLOv8-M, and YOLOv8-S variants at 5.70 FPS, 8.96 FPS, and 17.61 FPS, respectively, on full-resolution frames (Yu et al., 16 Jul 2025).

Other top-ranked teams exposed complementary design patterns. zhwa2003 used an intersection-based ensemble of two Cascade R-CNN + Swin Transformer + RFLA detectors to suppress false positives, especially around power poles, followed by Hybrid-SORT and interpolation. elsalabA combined Co-DETR, Copy-Paste augmentation, SAHI, and BoostTrack++, and reported a severe drop when BoostTrack++ was replaced with BoT-SORT. sgm used a six-output YOLOv7/YOLOv12 ensemble with confidence-based Adaptive Weighted Boxes Fusion and OC-SORT with DIoU cost. xmba15 combined a CenterNet ensemble with post-fusion false-positive rejection and frame-to-frame camera motion compensation via DISK, LightGlue, and affine transformation before OC-SORT. Taken together, these entries suggest that challenge success depended less on generic MOT recipes than on explicit accommodation of tiny targets, ego-motion, false-positive control, and association under low-overlap geometry (Kondo et al., 17 Jul 2025).

6. Significance, interpretive issues, and future directions

SMOT4SB established three technical pillars for subsequent work: a dedicated UAV bird-tracking dataset, a small-object-aware tracking metric, and a challenge ecosystem that made methods directly comparable under realistic motion entanglement. One immediate implication is evaluative rather than architectural: IoU-based HOTA can underestimate perceptual tracking quality in the tiny-object regime, whereas SO-HOTA credits close-but-non-overlapping predictions and therefore differentiates methods more meaningfully under small displacements (Kondo et al., 17 Jul 2025).

The challenge outcomes also weakened two common assumptions imported from larger-object MOT. First, strong appearance modeling is not necessarily central when most targets are only a few dozen pixels. The winning method was explicitly “completely independent of appearance information,” and several strong entries relied heavily on motion, geometry, or false-positive suppression rather than on classical Re-ID embeddings (Yu et al., 16 Jul 2025, Kondo et al., 17 Jul 2025). Second, naive full-frame training is not obviously optimal when the source imagery is 4K and the objects are tiny; training-time slicing, SAHI-style inference, Copy-Paste augmentation, and ensemble fusion emerged as recurrent responses to the same scale and memory constraints (Yu et al., 16 Jul 2025, Kondo et al., 17 Jul 2025).

The challenge papers identify several future directions: trajectory forecasting, confidence-aware tracking evaluation, better motion modeling under entangled motion, explicit modeling of flocking behavior through social interaction models, deployment on resource-limited platforms through quantization and compression, and cross-domain transfer to other small-object scenarios such as drones, insects, satellites, and small ground vehicles (Kondo et al., 17 Jul 2025, Yu et al., 16 Jul 2025). A related terminological caution is that “SMOT4SB” refers to small multi-object tracking for spotting birds and should not be conflated with the earlier generic tracker “SMOT: Single-Shot Multi Object Tracking,” which used an SSD-based tracking-by-re-detection design for other MOT benchmarks (Li et al., 2020).

In the broader literature on aerial wildlife vision, SMOT4SB occupies the transition point from static-frame bird detection toward trajectory-centric analysis. The predecessor SOD4SB challenge had already established that high-resolution UAV bird detection is dominated by tiny targets and severe localization sensitivity; SMOT4SB extends that regime into identity maintenance, where the central research question becomes how to preserve tracks when detection, overlap, and appearance are all intrinsically weak (Kondo et al., 2023, Kondo et al., 17 Jul 2025).