Researchers release AVID dataset: 68.1% of targets have a bounding box side of just 8–16 pixels

Researchers from the Georgia Institute of Technology, ARTPARK at the Indian Institute of Science, and Ramaiah Institute of Technology have introduced AVID, a dataset and robustness benchmark for detecting very small aerial targets. The set contains 32 040 frame pairs from 115 videos. In each pair, ground-based visible-band (RGB) and infrared (IR) cameras film the same scene, and the frame shows birds and drones.

For the 35 952 labeled targets, the authors computed the geometric mean of the bounding box width and height. For 68.1% of targets, it is just 8–16 pixels.

The cameras also view the scene slightly differently: in 16.3% of the 264 pairs checked, the offset between RGB and IR is larger than the median target side.

What is in the dataset

The footage comes from the Drone Detection Dataset, assembled by Svanström and colleagues. The bird–drone pairs were selected for the IEEE ICIP 2025 Video and Image Processing Cup. The contest's platform for identifying drone payloads is not part of AVID.

The annotations cover 35 952 objects: 15 471 birds and 20 481 drones. The boxes and classes are inherited from the contest organizers. The training split contains 28 790 pairs from 102 videos, and the validation split has 3 250 pairs from 13 other videos. In total, the set includes 76 videos with drones and 39 with birds. Clips run from 50 to 329 frame pairs, with a median of 306 and a mean of 278.6. Video IDs let researchers design evaluation so that neighboring frames from the same clip do not end up in both training and test.

The authors include an important caveat: an RGB–IR pair means the frames correspond to each other by scene, but they are not aligned pixel to pixel. The cameras have different optics, so a target can appear at different coordinates. The authors do not treat this misalignment as random noise; it is part of the evaluation conditions.

Tiny targets and camera offset

By the COCO criterion for small objects, 99.0% of targets count as small. That distribution is too uniform, so the authors break sizes into finer bins. The 8–16 pixel range alone contains 24 489 annotations.

For the offset between cameras, the authors introduce two tools: a relative alignment ratio and a "misalignment load curve." Both measure displacement in target sides rather than in pixels. A five-pixel offset barely affects a large object, but for a target ten pixels wide it is very large. In a sample of 300 pairs, the estimated global translation exceeds 5 pixels for 32.3%. Among 264 pairs with matching annotations, 16.3% have an offset larger than one median target side.

Interference test and separability analysis

The second part of the work is a controlled corruption test. It simulates seven families of image degradation at five severity levels, including noise, blur, lighting changes, and camera instability. The corruptions can be applied to RGB only, IR only, or both channels at once, so the affected sensor becomes an explicit experimental variable. The authors do not assume that the nominal strength of a corruption equals its actual difficulty for a detector.

The authors also ran an analysis in a fixed low-dimensional feature space. It shows how target size, corruptions, and added offsets between modalities affect the overlap between the "bird" and "drone" classes. The measure is the Bhattacharyya error bound: the lower the value, the better the classes separate. At a target size of 12–16 pixels, it is 0.148 for RGB, 0.198 for IR, and 0.046 for combined features. Combining sensors cuts the overlap substantially, even though IR on its own does worse than RGB here.

These values describe separability in the chosen features and density models, not the accuracy of a trained detector. The analysis also relies on the annotated target positions.

The authors propose using AVID for research that accounts for scale, alignment, and the characteristics of individual sensors. They say it helps identify the conditions under which RGB–IR fusion pays off and those under which it makes no sense. The preprint is published on arXiv in the cs.CV category.