Strict-Small Object Tracking

• Strict-small tracking is a framework for detecting and tracking extremely small objects in high-resolution imagery, addressing challenges like minimal appearance features and complex motion.
• SliceTrain enhances detection by partitioning images into overlapping tiles with stochastic augmentations, thereby increasing effective resolution and training diversity.
• Adaptive association leverages EMA-based velocity maintenance and bounding-box expansion to robustly manage erratic motion and occlusions, yielding superior tracking performance.

Strict-small tracking refers to frameworks and methodologies targeting the detection and tracking of extremely small, often agile, objects in visual input under highly challenging conditions. These systems must address the compounded issues of poor object appearance, complex scene and camera motion, pervasive occlusion, and frequent association ambiguities. Exemplified by benchmarks such as SMOT4SB, strict-small tracking pushes the boundaries of MOT (multi-object tracking) in domains such as aerial wildlife monitoring, sports analytics, and surveillance, where objects can be only a few pixels wide in high-resolution frames, rendering many standard detection and tracking assumptions invalid (Yu et al., 16 Jul 2025).

1. Challenges in Strict-Small Object Tracking

The fundamental obstacles in strict-small tracking are:

• Scarcity of appearance features: Small objects contribute minimal pixel area, preventing the use of rich, discriminative descriptors both for detection and for re-identification across frames. For example, tracking birds (~10x20 px in 4000x2000 px images) or players in wide-angle soccer sequences renders conventional deep features unreliable (Liu et al., 2024, Hurault et al., 2020).
• Entangled and irregular motion: The compounded dynamics of camera (especially UAV or moving sports cameras) and target objects result in complex motion paths, often with high acceleration and abrupt direction changes (Yu et al., 16 Jul 2025, Song et al., 2024).
• Frequent occlusions and dense target arrangements: Group behaviors, such as bird flocking or sports team formations, induce severe identity ambiguities and further aggravate association difficulties.
• Mismatched temporal and spatial cues: Standard frame resizing in detectors exacerbates the sub-pixel scale of targets, while naive sliding-windows yield excessive computation and false positives (Liu et al., 2024).

2. Detection Strategies: Slicing and Augmentation

Modern strict-small tracking pipelines emphasize adapting the detection stage to the characteristics of small-object regimes.

SliceTrain Framework

The YOLOv8-SMOT system introduces "SliceTrain," a training enhancement framework built on two pillars (Yu et al., 16 Jul 2025):

• Deterministic full-coverage slicing: High-resolution images are partitioned into overlapping tiles (examples: 1280x1280 px, 20% overlap). This ensures that each object is present at full or near-full pixel scale in at least one tile, maximizing effective resolution.
• Slice-level stochastic augmentation: Each tile undergoes independent, random augmentations including mosaic composition, color jittering, geometric transforms (scale ±20%, rotation ±15°, translation ±10%), and random flips. This produces high spatial and visual coverage per epoch.

This approach dramatically increases the effective batch size and diversity, enabling YOLO variants to learn small-scale detection more robustly.

3. Tracking Methodologies: Association and Motion Handling

Strict-small tracking is dominated by data association, which itself is uniquely challenging due to poor appearance cues and target ambiguity.

Adaptive Association and Velocity Maintenance

• Motion direction maintenance (EMA): YOLOv8-SMOT maintains target velocity using an exponential moving average (v_EMA) with decay (α=0.8), replacing instant estimates or pure Kalman-filter outputs. This stabilizes matching in the presence of irregular or abrupt trajectory changes by enforcing smoother velocity priors.
• Bounding-box expansion for similarity computation: Prior to matching, predicted and observed bounding boxes are uniformly expanded (scale factor s=2), and similarity is computed as a blend of IoU over these enlarged boxes and normalized center distance:

$$\text{Similarity} = \frac{\text{ExpandedIoU} - d_\text{norm} + 1}{2}$$

where $$d_\text{norm} = \|c_\text{pred} - c_\text{det}\|_2 / D_\text{max}$$ and $D_\text{max}$ is the diagonal length of the expanded box union.

• Multi-stage matching pipeline: The OC-SORT backbone is modified to perform successive rounds of association:
  1. High-confidence: using the adaptive similarity metric and EMA-stabilized motion.
  2. Rescue: for occluded or lost tracks, with thresholds progressively relaxed.
  3. Heuristic: further relaxation for final ID recovery.

Grid search is used to tune thresholds, e.g., IoU_thresh_primary=0.25 with decrement per stage ΔIoU=−0.08 (Yu et al., 16 Jul 2025).

4. Evaluation Protocols and Benchmark Results

Strict-small tracking frameworks are evaluated with metrics sensitive to both detection and association, such as SO-HOTA, SO-DetA, and SO-AssA (detection and association accuracy for small objects).

• YOLOv8-SMOT (Yu et al., 16 Jul 2025) achieves SO-HOTA=55.205 on the SMOT4SB public test set, compared to a baseline of 10.676. Ablations show the cumulative gain from EMA (+3.716), bounding-box expansion (+3.471), and distance penalty (+3.818).
• SliceTrain’s implementation uses 1280x1280 tile size, 20% overlap, batch size 6 per GPU (vs. 1 for full-image training), and augmentation probabilities 0.5 per transformation. Tracker hyperparameters include EMA decay α=0.8, expansion scale s=2, and primary IoU 0.25.

Model	SO-HOTA	SO-DetA	SO-AssA	Mem (MiB)	FPS
YOLOv8-L + SMOT	55.205	51.716	59.082	5036	5.70
YOLOv8-M + SMOT	54.426	49.529	59.962	3884	8.96
YOLOv8-S + SMOT	53.808	48.388	59.979	3190	17.61
Baseline	10.676	9.788	11.671	2346	8.88

These results demonstrate a step change in practical multi-object small-target tracking.

5. Architectural Integration and Implementation Considerations

• Detection and tracking integration: The tracking pipeline is detector-agnostic after preprocessing but performs best when the detector is explicitly optimized via the SliceTrain protocol for small targets.
• Scalability and throughput: The use of per-tile training, large batch sizes, and augmentation increases both GPU memory usage and throughput, compared to naive full-image training. The pipeline achieves up to 17.61 FPS (YOLOv8-S + SMOT), supporting near real-time deployment (Yu et al., 16 Jul 2025).
• Release and reproducibility: All code, hyperparameters, and pretrained weights are available at https://github.com/Salvatore-Love/YOLOv8-SMOT.

6. Impact and Domain Generalization

Strict-small tracking, via frameworks such as YOLOv8-SMOT, has enabled reliable multi-object tracking under the strictest appearance- and motion-constrained regimes encountered in aerial wildlife studies, sports analytics, and surveillance. Innovations in deterministic slicing, stochastic augmentation, velocity aggregation, and adaptive similarity have generalized well across datasets and target species, as independently validated in the MVA 2025 "Finding Birds" challenge (Yu et al., 16 Jul 2025). The result is a robust, detection-agnostic foundation for tracking regimes previously considered intractable.