—————— You can preview our dataset in the ‘Preview’ folder, which contains a subset of our data with resized resolution for demonstration purposes.

HINT: the original 4K-Resolution video can be downloaded here: https://pan.baidu.com/s/1dilMHXuHYVE15FQmH_8hLw?pwd=seek ——————

A2Seek: A Reasoning-Centric Benchmark for UAV Video Anomaly Understanding

Data Overview

The A2Seek dataset is the first comprehensively integrated large-scale aerial-view video anomaly detection dataset, uniquely tailored to real-world scenarios captured by unmanned aerial vehicles (UAVs). It comprises over 23 hours of untrimmed 4K video, distributed across 542 drone-shot sequences, spanning 10 diverse campus environments and 27 functional sub-scenes such as roads, playgrounds, rooftops, academic buildings, and open spaces. The dataset includes annotations for 20 types of anomaly behaviors, such as fighting, robbery, falling, playing with water, jaywalking, climbing over walls, and sneaking. These anomalies are further categorized based on risk severity into high risk, moderate risk, and low risk, and by temporal duration into long-term, short-term, and instantaneous anomalies.

Unlike conventional ground-based surveillance datasets, A2Seek leverages UAV-mounted cameras equipped with wide-angle, telephoto, and infrared lenses, capturing scenes at varying altitudes (10–60 meters) and speeds (0–20 m/s). The data encompasses a wide range of conditions — daytime, night, foggy weather, rainy days, and twilight — enabling rich diversity in visual contexts and environmental dynamics. The annotation system in A2Seek goes beyond simple labels, offering multi-level semantic reasoning annotations: frame-level timestamps, region-level bounding boxes, anomaly category tags, and most notably, natural language causal explanations. These annotations empower models not only to detect “where” anomalies occur but also to reason about “why” they are anomalous.

A unique feature of A2Seek is its integration of the Graph-of-Thought (GoT) structure, guiding models through progressive reasoning stages such as Trigger, Diagnose, Reasoning, Reflection, and Seeking. Particularly, the “Seeking” stage introduces a novel mechanism where the model is encouraged to identify areas that require additional inspection, thereby simulating UAV-like motion behavior to dynamically focus on informative regions. This makes A2Seek especially suited for developing models that can handle region-aware anomalies, which are often spatially sparse and context-dependent.

For more information, check this page

Data Collection

• introduction
  • 
  • The data collection process of the A2Seek dataset was meticulously designed to maximize real-world diversity, environmental realism, and dynamic scene coverage, providing a robust foundation for reasoning-centric aerial anomaly detection.
  • For platform and sensors, the data was collected using a DJI M30T industrial drone equipped with a tri-sensor setup: a wide-angle camera, a telephoto camera, and a 640×512 thermal infrared imager. This combination ensures robust visual coverage across varied spatial scales and lighting conditions, including low-light or nighttime scenarios. In addition, a 50 Hz GNSS/IMU system onboard continuously logs precise position and motion data of the UAV, supporting accurate spatiotemporal alignment and future tasks like trajectory-aware modeling or contextual anomaly understanding.
  • In terms of flight parameters, the UAV flew at altitudes ranging from 10 to 60 meters and speeds between 0 and 20 m/s, simulating real-world deployment scenarios such as urban patrol, perimeter monitoring, or emergency surveillance. The flight modes covered a wide variety of UAV behaviors, including hovering, straight-line cruising, circular orbiting, and area-based grid scanning, allowing for diverse visual angles and spatiotemporal continuity in the captured footage.
  • Regarding scenario diversity, A2Seek was built from 542 untrimmed 4K@30 FPS videos, recorded across 10 representative campus environments and subdivided into 27 distinct sub-scenes, such as roadways, playgrounds, rooftops, entrances, and building courtyards. The full dataset spans over 23 hours of footage, including 19.3 hours of normal activity and 3.79 hours of complex anomalous behaviors, encompassing both frequent and rare events.
  • To enhance sample utility and annotation precision, the team curated over 42,000 keyframes along with their corresponding video segments. These samples span more than 20 anomaly categories—such as falling, fighting, water-related play, and red-light violations—and are distributed across four seasons, multiple times of day (morning, midday, night), and varied weather conditions including sunny, cloudy, foggy, and rainy environments. This rich environmental variation ensures that A2Seek provides high value for training and evaluating models under realistic, complex, and distribution-shifting conditions.
• Platform & Sensors
  DJI M30T drone equipped with a wide-angle, telephoto, and 640×512 thermal imaging camera; onboard GNSS/IMU logging at 50 Hz.
• Flight Parameters
  • Altitude: 10–60 m
  • Speed: 0–20 m/s
  • Flight Modes: hover, straight-line patrol, orbit, and area scan
• Scenarios
  Captured 542 4K@30 FPS videos across 10 campus environments and 27 sub-scenes, totaling over 23 hours (19.3 h normal, 3.79 h anomalous).
• Sample Scale
  • Over 42,000 carefully selected keyframes with corresponding video clips
  • More than 20 anomaly categories (e.g., fall, fight, water play, red-light violation) spanning seasons, times of day, and weather conditions (sunny, cloudy, foggy, rainy)

Data Annotation

• Multi-Stage Annotation Workflow
  1. Stage 1: Temporal Annotation: Annotators use a temporal labeling tool to scan drone videos and identify the start and end timestamps of each anomalous behavior. Each detected anomaly is assigned a category label (e.g., "Running") and stored in a structured JSON format with temporal boundaries.
  2. Stage 2: Single Frame Spatial Annotation & Explanation: For each temporal anomaly segment, a representative key frame is selected. Annotators manually draw bounding boxes around the anomalous regions and provide natural language descriptions (e.g., "A person is running on the sidewalk, which is against the regulations.") to add semantic context.
  3. Stage 3: Spatial Label Propagation & Verification: Using a pretrained tracking model, the single-frame spatial annotation is propagated across all relevant frames. A validation module then checks the quality of these labels. Unqualified outputs are flagged for review and correction, ensuring consistency and accuracy of full-frame spatial annotations.
  4. Stage 4: Label Merging with COT Reasoning: Multimodal Vision-Language Models (VLMs) are introduced at this stage. With spatial, temporal, and instructional inputs, the models generate structured Chain-of-Thought (CoT) reasoning outputs. These are merged with prior annotations to produce unified, explainable, and verifiable final labels.

---