AVID: A Weakly Aligned RGB-Infrared Dataset and Perturbation Benchmark for Tiny Aerial Targets

Dharini Raghavan*, Vishnu Rao*, Sethu Selvi S
* The first two authors contributed equally and share first authorship.

Paper: Final preprint PDF · Source LaTeX · Code and reproducibility CSVs · Research website

Dataset overview

AVID contains 32,040 corresponding RGB-infrared frame pairs from 115 ground-based surveillance videos, with 35,952 existing bird/drone bounding-box annotations. It was prepared from the bird-drone data previously curated for the IEEE ICIP 2025 Video and Image Processing Cup, using original footage from the Drone Detection Dataset collected by Svanström, Alonso-Fernandez, and Englund.

The RGB and infrared streams correspond to the same scene, but the cameras are not necessarily pixel-aligned. The original videos and existing annotations were not newly captured or labeled for this paper.

Quantity	Train	Validation	Total
RGB-IR pairs	28,790	3,250	32,040
Videos	102	13	115
Bird instances	14,113	1,358	15,471
Drone instances	18,335	2,146	20,481
All instances	32,448	3,504	35,952

Frames are 320 × 256 pixels; 68.1% of annotated boxes have geometric-mean sides of 8–16 pixels and 99.0% are COCO-small. Classes are 0=bird, 1=drone in the provided YOLO-style labels.

Download

python -m pip install --upgrade "huggingface_hub[cli]"
hf download Dharini24/AVID --repo-type dataset --local-dir ./AVID

AVID-C and alignment

AVID-C applies seven types of simulated image degradation at five severity levels to RGB, infrared, or both modalities. The released implementation and parameters reproduce the transformations without publishing separate corrupted copies of the images.

Global RGB-infrared translations were estimated on 300 sampled pairs (median shift 1 px; 32.3% exceed 5 px). Target-relative alignment analysis uses the 264 pairs with matched annotations, of which 16.3% exceed one median target side. The unit-scale alignment burden for the 264 pairs is 0.265. These are estimated global translations, not ground-truth pixel correspondence for all frames.

The paper studies Bhattacharyya and Chernoff upper bounds and a KDE comparison using a fixed PCA representation. These values represent fitted class separation, not detector error rates. Evaluation on other cameras, environments and models is needed for deployment claims.

Source attribution and licensing

The original Drone Detection Dataset was collected and publicly shared by Fredrik Svanström, Fernando Alonso-Fernandez and Cristofer Englund. Source repository: https://github.com/DroneDetectionThesis/Drone-detection-dataset. Please cite Svanström et al., A Dataset for Multi-Sensor Drone Detection, Data in Brief 39 (2021), 107521, https://doi.org/10.1016/j.dib.2021.107521.

License clarification: the source imagery was released under CC0 1.0 Universal. The AVID dataset-package listing uses CC BY-NC 4.0, which does not change the original CC0 status of the underlying source media. Code in the GitHub repository uses MIT.

Responsible use and limitations

The dataset supports research on tiny-target detection, weak alignment and image degradation. It does not provide person-identification or personal-attribute labels. UAV surveillance can have privacy and dual-use implications. The label review history is incomplete, alignment was measured on a sample, and simulated degradation does not cover every real-world failure mode. These limitations are described in the paper.

Temporary citation (update after arXiv/DOI assignment)

@misc{raghavan2026avid,
  title = {AVID: A Weakly Aligned RGB-Infrared Dataset and Perturbation Benchmark for Tiny Aerial Targets},
  author = {Raghavan, Dharini and Rao, Vishnu and S, Sethu Selvi},
  year = {2026},
  note = {Manuscript and dataset release; arXiv ID and DOI pending},
  url = {https://github.com/rdharini2001/ICIP_2025_Data_Release}
}