OD-VIRAT Tiny Object Detection Benchmark

• OD-VIRAT Tiny is a surveillance-oriented benchmark derived from VIRAT Ground 2.0, focusing on high camera angles, severe occlusion, and cluttered backgrounds.
• It comprises 19,860 images and 288,901 annotations across five classes, offering a challenging and realistic testbed for small-object detection.
• Baseline results reveal that transformer-based methods with deformable attention excel in localizing tiny, occluded objects compared to one-stage detectors.

Searching arXiv for the OD-VIRAT paper and a few directly related detector papers to ground the article. OD-VIRAT Tiny is the smaller of two visual object detection benchmarks introduced in “OD-VIRAT: A Large-Scale Benchmark for Object Detection in Realistic Surveillance Environments” (Ullah et al., 16 Jul 2025). It is a surveillance-oriented object detection benchmark derived from the VIRAT Ground 2.0 corpus and designed to train and evaluate detectors under conditions characterized by high camera angles, large scene coverage, complex backgrounds, heavy occlusion, and very small objects. The benchmark comprises 19,860 images and 288,901 annotated bounding-box instances across five object classes—Person, Car, Vehicle, Bike/Bicycle, and Carrying_object—and is distributed in COCO-style JSON with axis-aligned boxes represented as $[x, y, width, height]$ in pixel units (Ullah et al., 16 Jul 2025).

1. Benchmark definition and scope

OD-VIRAT Tiny is described as the “light” version of the broader OD-VIRAT benchmark, with the explicit purpose of keeping a realistic surveillance detection task manageable while preserving its principal difficulties (Ullah et al., 16 Jul 2025). The benchmark is built from video material recorded in ten distinct outdoor scenes, including construction sites, parking lots, streets, and open areas. These scenes are captured by stationary cameras mounted at significant height, often on rooftops, and recorded at resolutions of $1280 \times 720$ or $1920 \times 1080$ (Ullah et al., 16 Jul 2025).

The defining characteristic of the benchmark is its emphasis on realistic surveillance geometry. Objects may appear at distances up to several tens of meters, and many targets occupy only a few dozen pixels. The benchmark therefore stresses detector behavior in a regime where scale, occlusion, and scene clutter are not incidental nuisances but structural properties of the data (Ullah et al., 16 Jul 2025). This suggests that OD-VIRAT Tiny is intended not merely as a reduced dataset, but as a compact stress test for small-object detection in overhead or elevated-view surveillance imagery.

The benchmark uses five object categories. Person and vehicle-centric classes dominate the semantic space, while Carrying_object introduces human-interaction context through examples such as suitcases and carts (Ullah et al., 16 Jul 2025). The paper does not publish a per-class breakdown, and that omission is itself methodologically relevant: analyses of class imbalance beyond the aggregate instance count cannot be made from the published summary.

2. Construction from VIRAT Ground 2.0

OD-VIRAT Tiny is derived from the VIRAT Ground 2.0 corpus by filtering source videos for annotation consistency. From an original set of 329 videos, 260 were retained after sanity checks on event-to-object mapping (Ullah et al., 16 Jul 2025). The retained videos were then converted to frames with a 30-frame skip, approximately 1 frame per second, in order to control redundancy and keep the tiny variant manageable (Ullah et al., 16 Jul 2025).

This construction procedure yields 19,860 total images and 288,901 annotated instances (Ullah et al., 16 Jul 2025). The annotation format follows the COCO schema, with JSON files organized around images[], annotations[], and categories[], and category IDs ranging from 1 to 5 (Ullah et al., 16 Jul 2025). Each annotation consists of a single axis-aligned bounding box. The use of COCO-style representation is technically consequential because it permits direct reuse of standard detection toolchains, evaluation code, and metric definitions without introducing a dataset-specific annotation parser.

The dataset’s scene and camera properties define its difficulty profile. Heavy background clutter is reported, including building facades and vegetation; frequent inter-object occlusions are present; and object scales range from less than 32 pixels to more than 300 pixels (Ullah et al., 16 Jul 2025). A plausible implication is that OD-VIRAT Tiny is particularly useful for evaluating the failure modes of detectors whose feature hierarchies or assignment mechanisms are sensitive to aggressive downsampling and crowded scenes.

3. Split design and data statistics

The benchmark is partitioned at the video level into train, validation, and test subsets. The split assigns 156 videos to training, 52 to validation, and 52 to testing (Ullah et al., 16 Jul 2025). In frame counts, this corresponds to 12,501 training images, 4,573 validation images, and 2,786 test images, or 62.94%, 23.03%, and 14.02% respectively (Ullah et al., 16 Jul 2025).

Split	Frames	Percentage
Train	12,501	62.94%
Val	4,573	23.03%
Test	2,786	14.02%

Object-scale statistics are described with reference to COCO definitions. Small objects are defined as having area $< 32^2$, medium objects as $32^2 \leq \text{area} < 96^2$, and large objects as $\text{area} \geq 96^2$ (Ullah et al., 16 Jul 2025). Although exact size histograms are not tabulated, the authors report that a significant fraction, greater than 50%, of all objects are small, and that moderate to heavy occlusions appear in over 30% of frames (Ullah et al., 16 Jul 2025).

These statistics place OD-VIRAT Tiny in a regime substantially more demanding than many conventional object detection benchmarks dominated by medium and large instances. The predominance of small objects has direct implications for anchor design, feature pyramid resolution, query efficiency in transformer detectors, and NMS behavior in crowded scenes. The reported prevalence of occlusion also makes the dataset relevant for studying partial visibility and inter-instance ambiguity under standard axis-aligned supervision.

4. Evaluation protocol

All models are evaluated on the test split using standard COCO-style detection metrics (Ullah et al., 16 Jul 2025). The benchmark adopts Intersection over Union as the localization criterion:

$$\mathrm{IoU}(B_p, B_{gt}) = \frac{\lvert B_p \cap B_{gt}\rvert}{\lvert B_p \cup B_{gt}\rvert}.$$

Average Precision and mean Average Precision are defined as:

$$\mathrm{AP} = \int_{0}^{1} p(r)\,dr, \quad \mathrm{mAP} = \frac{1}{C}\sum_{c=1}^{C}\mathrm{AP}_c,$$

with $C=5$ classes (Ullah et al., 16 Jul 2025). The reported metrics include $mAP_{50}$ at IoU $1280 \times 720$0, $1280 \times 720$1 at IoU $1280 \times 720$2, and the COCO-style averaged $1280 \times 720$3 in increments of 0.05, reported simply as “mAP” in the benchmark tables (Ullah et al., 16 Jul 2025). Scale-wise metrics are also reported as $1280 \times 720$4, $1280 \times 720$5, and $1280 \times 720$6 for small, medium, and large objects (Ullah et al., 16 Jul 2025).

The choice of COCO-style averaging is significant in this setting because high $1280 \times 720$7 alone can obscure localization instability on small or partially occluded objects. By including $1280 \times 720$8 and $1280 \times 720$9, the benchmark differentiates detectors that merely produce coarse overlaps from those that maintain box quality under stricter IoU thresholds. In surveillance imagery, where boxes can span only a few dozen pixels, even small coordinate errors can produce substantial metric degradation, making multi-threshold evaluation especially informative.

5. Baseline detector results

Five state-of-the-art detectors were benchmarked on OD-VIRAT Tiny: RTMDET, YOLOX, RetinaNet, DETR, and Deformable-DETR (Ullah et al., 16 Jul 2025). The reported comparison reproduces the best pretrained variant of each family on the test set (Ullah et al., 16 Jul 2025).

Model	Backbone	Key results
RTMDET Large	CSPNeXt-P5	mAP 59.2, $1920 \times 1080$0 73.1, $1920 \times 1080$1 63.4, $1920 \times 1080$2 30.8, $1920 \times 1080$3 73.7, $1920 \times 1080$4 92.0
YOLOX Large	CSPDarkNet	mAP 52.9, $1920 \times 1080$5 78.7, $1920 \times 1080$6 59.0, $1920 \times 1080$7 36.1, $1920 \times 1080$8 63.9, $1920 \times 1080$9 80.0
RetinaNet (ResNet101)	ResNet101	mAP 70.8, $< 32^2$0 85.5, $< 32^2$1 71.6, $< 32^2$2 59.4, $< 32^2$3 88.3, $< 32^2$4 96.7
DETR (ResNet50)	ResNet50	mAP 56.8, $< 32^2$5 72.4, $< 32^2$6 60.0, $< 32^2$7 28.3, $< 32^2$8 73.9, $< 32^2$9 91.8
Deformable-DETR ++	ResNet50	mAP 75.0, $32^2 \leq \text{area} < 96^2$0 93.3, $32^2 \leq \text{area} < 96^2$1 76.3, $32^2 \leq \text{area} < 96^2$2 64.9, $32^2 \leq \text{area} < 96^2$3 88.2, $32^2 \leq \text{area} < 96^2$4 97.5

Among the listed models, Deformable-DETR ++ achieves the highest overall test performance, with $32^2 \leq \text{area} < 96^2$5, $32^2 \leq \text{area} < 96^2$6, and $32^2 \leq \text{area} < 96^2$7 (Ullah et al., 16 Jul 2025). RetinaNet with ResNet101 follows with $32^2 \leq \text{area} < 96^2$8, $32^2 \leq \text{area} < 96^2$9, and overall mAP 70.8 (Ullah et al., 16 Jul 2025). On small objects specifically, Deformable-DETR ++ attains $\text{area} \geq 96^2$0, exceeding RetinaNet’s $\text{area} \geq 96^2$1, while both substantially exceed RTMDET Large and DETR on the same scale subset (Ullah et al., 16 Jul 2025).

The paper attributes Deformable-DETR’s advantage to deformable attention, stating that it excels at localizing small, occluded objects in cluttered backgrounds (Ullah et al., 16 Jul 2025). RetinaNet is characterized as a strong second-place detector, particularly on medium and large targets, with its FPN and focal-loss cited as contributing factors (Ullah et al., 16 Jul 2025). DETR is reported to underperform on very small objects and to converge more slowly, although it still benefits from end-to-end training (Ullah et al., 16 Jul 2025). YOLOX and RTMDET, while very fast, are described as struggling with the tiny, heavily occluded objects that dominate the benchmark (Ullah et al., 16 Jul 2025).

These outcomes are notable because the evaluated model families span one-stage, anchor-based, anchor-free, and transformer-based paradigms. The reported ordering suggests that, on OD-VIRAT Tiny, mechanisms that preserve or selectively attend to sparse fine-scale evidence confer a measurable advantage over architectures optimized primarily for real-time throughput.

6. Difficulty profile, interpretation, and methodological significance

The benchmark’s main difficulties are identified as small targets, often less than $\text{area} \geq 96^2$2 pixels, high camera elevations, complex and dynamic backgrounds, frequent inter-object occlusions, and severe scale variation (Ullah et al., 16 Jul 2025). These factors interact rather than appearing independently. For example, a small object partially occluded against cluttered foliage or facade texture creates simultaneous challenges in feature extraction, assignment, and classification calibration.

A common misconception in surveillance detection is that higher image resolution alone resolves small-object limitations. The benchmark description does not support that simplification. Although recordings are made at $\text{area} \geq 96^2$3 or $\text{area} \geq 96^2$4, many targets still occupy only a few dozen pixels because of long standoff distance and elevated viewpoint (Ullah et al., 16 Jul 2025). This indicates that nominal frame resolution should not be conflated with effective target resolution.

Another potential misunderstanding is that OD-VIRAT Tiny is merely a reduced-size convenience subset. Its construction choices instead preserve difficult scene geometry and annotation density while limiting redundancy through frame sampling at approximately 1 Hz (Ullah et al., 16 Jul 2025). A plausible implication is that the “Tiny” designation refers primarily to benchmark scale relative to OD-VIRAT Large, rather than to a simplification of the detection problem.

Within the broader context of object detection research, OD-VIRAT Tiny functions as a benchmark for realistic surveillance imagery rather than generic natural-image detection. The paper explicitly positions its experimental study as the first examination of recently published state-of-the-art object detection architectures on this object detection-specific variant of the VIRAT dataset under challenging conditions such as complex backgrounds, occluded objects, and small-scale objects (Ullah et al., 16 Jul 2025). Its methodological significance therefore lies in exposing how detector families behave when the operating regime departs from the assumptions implicit in many standard benchmarks.

7. Reported directions for improvement

The benchmark study includes explicit recommendations for future improvements (Ullah et al., 16 Jul 2025). First, it recommends integrating multi-scale deformable attention in one-stage and two-stage detectors to strengthen small-object recall. Second, it recommends incorporating explicit occlusion-aware modules or adversarial occlusion augmentation during training. Third, it recommends fusing temporal context from adjacent frames to improve detection of partially visible objects. Fourth, it recommends designing lightweight, high-resolution feature extractors that preserve sub-pixel structure (Ullah et al., 16 Jul 2025).

These recommendations are framed by the observed behavior of the baseline models. Since Deformable-DETR leads the benchmark and is specifically described as excelling in localizing small, occluded objects in cluttered backgrounds, the proposed emphasis on deformable attention is directly tied to the empirical ranking rather than being presented as a generic architectural preference (Ullah et al., 16 Jul 2025). Likewise, the call for temporal fusion reflects the surveillance setting’s video provenance, even though the benchmark itself is formulated as an image-based object detection task.

In summary, OD-VIRAT Tiny is a compact but difficult benchmark for object detection in realistic surveillance environments. Its 19,860 images, 288,901 annotations, five-category label space, video-level train/validation/test split, and COCO-style evaluation protocol combine to create a benchmark in which small-object recall, occlusion robustness, and localization precision are central determinants of performance (Ullah et al., 16 Jul 2025). The baseline results indicate that transformer-based methods with deformable attention currently perform best on this task, while one-stage and real-time-oriented detectors remain more constrained by the benchmark’s dominant regime of tiny and heavily occluded targets (Ullah et al., 16 Jul 2025).