OD-VIRAT Large Surveillance Benchmark

• OD-VIRAT Large is a surveillance-specific object detection benchmark derived from VIRAT Ground 2.0 with COCO-style annotations for five classes.
• It comprises 599,996 images and 8.7 million instances across 10 distinct surveillance scenes, emphasizing challenges like occlusion and scale variation.
• The dataset employs a 0 frame-skip strategy and detailed annotation pipeline, enabling robust evaluation of detectors on small, distant objects.

Searching arXiv for the benchmark paper and closely related VIRAT context papers. tool call: arxiv_search {"query":"OD-VIRAT Large benchmark realistic surveillance environments VIRAT Ground 2.0", "max_results": 5, "sort_by": "relevance"} tool call: arxiv_search {"query":"(Ullah et al., 16 Jul 2025)", "max_results": 10, "sort_by": "relevance"} OD-VIRAT Large is the large-scale object-detection version of the VIRAT Ground 2.0 surveillance dataset introduced in “OD-VIRAT: A Large-Scale Benchmark for Object Detection in Realistic Surveillance Environments” (Ullah et al., 16 Jul 2025). It is designed as a realistic surveillance benchmark for object detection rather than a generic natural-image corpus, and it targets the specific operating regime of fixed cameras mounted at significant height and distance, with wide-area coverage, complex backgrounds, frequent occlusion, and strong scale variation. In the benchmark design presented in the paper, OD-VIRAT Large is the full-frame, full-scale variant of OD-VIRAT, built from VIRAT Ground 2.0 by retaining every frame from the selected videos and converting the source annotations into COCO JSON bounding-box annotations for five classes (Ullah et al., 16 Jul 2025).

1. Dataset provenance and construction rationale

OD-VIRAT Large is derived from VIRAT Ground 2.0, which was originally an event-recognition dataset comprising 329 videos recorded by stationary ground cameras mounted at significant heights, mostly on top of buildings, across 11 scenes. The original scenes include construction sites, parking lots, streets, and open outdoor areas, and the source videos were recorded at resolutions of 1920×1080 or 1280×720 with frame rates around 25–30 FPS. For OD-VIRAT, 260 of the 329 videos were selected, while 69 videos were excluded because of incorrect temporal relations between objects and annotations; the retained videos cover 10 distinct scenes (Ullah et al., 16 Jul 2025).

The rationale for introducing OD-VIRAT Large is explicitly surveillance-specific. The benchmark is motivated by the claim that widely used detection datasets such as PASCAL VOC, COCO, Objects365, Open Images, and BDD100K do not reproduce the visual conditions of practical surveillance footage. The paper identifies the core mismatch as a combination of high camera elevation, long stand-off distance, wide-area observation, tiny or small objects, complex backgrounds, frequent occlusion, and large scale variation. This positions OD-VIRAT Large as a benchmark for studying detector behavior under a domain in which modern architectures may behave differently than they do on standard natural-image datasets.

A central design choice is that OD-VIRAT Large is formed with a 0 frame-skip strategy, meaning that every frame of each retained video is used. The resulting benchmark is therefore not a sparse sampling of surveillance frames, but the full-scale detection variant intended to serve as a “gigantic object detection benchmark dataset” for realistic surveillance imagery.

2. Scale, splits, and label space

OD-VIRAT Large contains 599,996 images and 8.7 million annotated instances drawn from 260 videos spanning 10 distinct surveillance scenes. Its fixed split is 156 videos for training, 52 for validation, and 52 for testing. At the image level, the split is 377,686 training images, 137,971 validation images, and 84,339 test images. The benchmark uses five object classes: Bike/Bicycle, Car, Carrying_object, Person, and Vehicle (Ullah et al., 16 Jul 2025).

The paper repeatedly contrasts OD-VIRAT Large with OD-VIRAT Tiny, which differs only in frame extraction strategy. Large uses 0 frame-skip, whereas Tiny uses 30 frame-skip, i.e. one frame every 30 frames. The class set and scenes are the same, but Tiny is roughly 30× smaller and was used for the experimental study because of computational constraints.

Property	OD-VIRAT Large	OD-VIRAT Tiny
Frame extraction	0 frame-skip	30 frame-skip
Images	599,996	19,860
Annotated instances	8.7 million	288,901

The split geometry is also noteworthy. Train and validation each contain 10 scenes, whereas test contains 8 scenes. The paper does not present this as a cross-scene generalization protocol or challenge-server setup; it is simply the standard benchmark partition. A plausible implication is that OD-VIRAT Large should be interpreted primarily as a fixed benchmark split rather than as a formally designed open-set scene-transfer benchmark.

3. Surveillance regime and visual difficulty

OD-VIRAT Large is explicitly organized around realistic daytime surveillance. The scene and acquisition characteristics emphasized in the paper are daytime scenes, stationary cameras, significant camera height, significant camera-to-object distance, wide-area observation, complex backgrounds, occluded objects, tiny or small objects, and objects of varying sizes. These properties are not incidental; they define the benchmark’s intended difficulty regime (Ullah et al., 16 Jul 2025).

This surveillance regime distinguishes OD-VIRAT Large from several adjacent benchmark families. Unlike COCO, PASCAL VOC, and Open Images, it is not composed of independently curated natural images with relatively prominent target instances. Unlike BDD100K, it is not ego-vehicle driving data. Unlike VisDrone, it is not aerial or drone imagery; the paper stresses that OD-VIRAT Large is captured by fixed ground surveillance cameras mounted high above the scene. The viewpoint is therefore elevated and distant, but still ground-based and stationary.

The benchmark’s difficulty is concentrated in three interacting factors. First, many targets are small or tiny because the cameras monitor large outdoor areas from long range. Second, occlusion is common, which degrades instance visibility even when targets are present. Third, cluttered backgrounds and strong scale variation make localization and classification harder than in close-up imagery. The paper also notes that numbers of bounding boxes per frame vary across scenes and splits, indicating nonuniform object density, although it does not define a formal crowding statistic.

A common misconception is that OD-VIRAT Large should be grouped with general-purpose large-scale object-detection datasets solely on the basis of image count. The paper instead frames its significance as domain-specific scale: while some earlier datasets are larger in absolute image count, OD-VIRAT Large is unusually large for a surveillance-specific object detection benchmark.

4. Annotation pipeline, representation, and evaluation protocol

The source VIRAT dataset provides three annotation files per video: events.txt for temporal event annotations, objects.txt for moving-object bounding boxes, and mapping.txt for linking bounding boxes to events. OD-VIRAT construction proceeds in three phases: retrieving frame-level bounding boxes from objects.txt, compiling the retrieved coordinates by associating them with frames, and converting the annotations from .txt to COCO .json format. The frame-level compilation step explicitly handles frames containing zero, one, or multiple objects, and the converted train/validation/test annotations are visually verified against the frames (Ullah et al., 16 Jul 2025).

As a result, OD-VIRAT Large provides bounding box annotations and class labels in COCO JSON format. The paper does not describe segmentation masks, keypoints, or tracking IDs as part of the OD-VIRAT detection annotations. This matters for task definition: OD-VIRAT Large is a pure object-detection benchmark in the presentation of the paper, even though its source corpus originated in event recognition.

The evaluation methodology follows COCO-style mean Average Precision. Reported metrics are $mAP$, meaning $mAP_{50:95}$, together with $mAP_{50}$, $mAP_{75}$, and the scale-stratified scores $mAP_S$, $mAP_M$, and $mAP_L$. The paper also reports Mean Performance under Corruption (mPC) in its robustness analysis, although it does not provide a formal metric equation in the provided text. This metric suite is conventional for object detection, but its use here is specifically intended to expose performance on small and distant surveillance targets.

5. Relationship to detector benchmarking

An important distinction in the paper is between benchmark construction and benchmarked experiments. OD-VIRAT Large is introduced as the full-scale dataset, but the reported detector experiments were conducted only on OD-VIRAT Tiny because of limited resources and time. Consequently, OD-VIRAT Large is accompanied by full dataset statistics, annotation methodology, and benchmark protocol, but not by a full large-scale leaderboard trained directly on the Large split (Ullah et al., 16 Jul 2025).

The experimental suite on Tiny includes RTMDet, YOLOX, RetinaNet, DETR, and Deformable-DETR. The tested variants include RTMDet Tiny, Small, Medium, and Large with CSPNeXt-P5; YOLOX Nano, Tiny, Small, Medium, and Large with CSPDarkNet; RetinaNet with ResNet18, ResNet50, and ResNet101; DETR with ResNet18, ResNet50, and ResNet101; and Deformable-DETR with PVT-Tiny or ResNet50, including the base model, iterative bounding box refinement, and two-stage Deformable-DETR. The training implementation uses PyTorch 2.1.0 and MMDetection, with 50 epochs for all models, and the paper reports results both with and without pre-trained weights.

The strongest quantitative result reported on OD-VIRAT Tiny is obtained by two-stage Deformable-DETR with ResNet50 and pretraining, reaching $mAP = 75.0$, $mAP_{50} = 93.3$, $mAP_{75} = 76.3$, $mAP_{50:95}$0, $mAP_{50:95}$1, and $mAP_{50:95}$2. Deformable-DETR with iterative refinement and ResNet50 with pretraining reaches $mAP_{50:95}$3, RetinaNet with ResNet101 and pretraining reaches $mAP_{50:95}$4, and Deformable-DETR with ResNet50 and pretraining reaches $mAP_{50:95}$5. The paper interprets these results as evidence that Deformable-DETR performs best on realistic surveillance imagery, RetinaNet is consistently second best and unexpectedly competitive, and RTMDet underperforms substantially under this regime.

The scale-sensitive results are especially relevant because OD-VIRAT’s central challenge is small and distant objects. Best reported $mAP_{50:95}$6 values with pretraining include 64.9 for two-stage Deformable-DETR with ResNet50, 64.5 for Deformable-DETR with iterative refinement and ResNet50, 60.3 for RetinaNet with ResNet50, and 59.4 for RetinaNet with ResNet101. By contrast, RTMDet Large with pretraining reaches $mAP_{50:95}$7 and $mAP_{50:95}$8, while YOLOX Large with pretraining reaches $mAP_{50:95}$9 and $mAP_{50}$0. The paper’s qualitative analysis further states that RetinaNet, YOLOX, and RTMDet can produce redundant bounding boxes, whereas DETR and Deformable-DETR produce cleaner and more precise detections.

The speed–accuracy trade-off on Tiny is also reported. YOLOX is the fastest family at 256.6 FPS on average; RetinaNet averages 149.8 FPS, DETR 148.4 FPS, Deformable-DETR about 145 FPS, and RTMDet 81.6 FPS. This suggests a practical division of roles: YOLOX is most attractive when real-time throughput is the primary constraint, whereas Deformable-DETR is preferred when surveillance-specific accuracy, especially on small objects, is the primary criterion.

6. Interpretation, limitations, and research uses

OD-VIRAT Large fills a specific gap between generic object-detection benchmarks and real deployment conditions in security and public-safety systems. Its significance lies in combining surveillance-specific imagery, COCO-format annotations, explicit train/validation/test partitions, and full-scale temporal retention from video into a single object-detection benchmark. Because the source videos come from stationary cameras mounted at significant height and distance, the dataset exposes long-range detection, occlusion, clutter, and scale variation as first-order rather than incidental challenges (Ullah et al., 16 Jul 2025).

Several limitations are also explicit. First, the paper does not provide full benchmarking results on OD-VIRAT Large; all detector experiments are on Tiny. Second, it does not provide an explicit class-frequency table for Large, so class imbalance is discussed only indirectly. Third, it does not formalize the metric equations in the provided text. Fourth, the test split contains 8 scenes while train and validation each contain 10, so scene coverage is not fully symmetric. Fifth, the data appears to be daytime only, limiting conclusions about night or low-light surveillance.

These limitations clarify how OD-VIRAT Large should and should not be interpreted. It is not yet a fully saturated benchmark with exhaustive large-scale baselines, nor is it a multimodal or multitask surveillance corpus in the form described here. Rather, it is a large surveillance detection dataset whose primary immediate contribution is to make realistic fixed-camera object detection experimentally accessible at scale. A common misunderstanding is that the strong Tiny results already constitute definitive evidence for Large-scale rankings. The paper does not make that claim; instead, Tiny functions as a computationally manageable proxy for the same surveillance conditions.

Within that framing, OD-VIRAT Large is suited to training surveillance-specific object detectors, studying small-object detection, evaluating occlusion robustness, benchmarking cross-scene generalization, pretraining or fine-tuning on realistic surveillance data, and developing methods for long-distance detection, background-clutter handling, and robustness to environmental degradation. The robustness experiments on Tiny—covering Gaussian Noise, Motion Blur, Snow, and Elastic Transform at severity levels 1 to 5—suggest that these use cases are not peripheral. They indicate that the OD-VIRAT regime is challenging not only in clean imagery, but also under perturbations that are operationally relevant in surveillance footage.

In that sense, OD-VIRAT Large is best understood as a domain-specific large-scale benchmark whose central contribution is not merely dataset size, but the codification of a surveillance operating regime in which fixed elevated viewpoints, distant targets, occlusion, clutter, and strong scale variation become the primary determinants of detector behavior.