Stream Contracts: An Auditable and Reproducible Evaluation Framework for RGB–Depth AAV Detection

Ekici S., Uyar M., Ustundag M.

IEEE ACCESS, cilt.14, ss.152458-152470, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 14
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1109/access.2026.3738662
  • Dergi Adı: IEEE ACCESS
  • Derginin Tarandığı İndeksler: Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, Directory of Open Access Journals
  • Sayfa Sayıları: ss.152458-152470
  • Bursa Uludağ Üniversitesi Adresli: Evet

Özet

Multimodal RGB–depth AAV detection studies commonly report comparative performance without explicitly specifying which image pairs are co-observable, how annotation ownership is enforced across modalities, or how pairing constraints alter the effective evaluation set. This lack of protocol transparency complicates reproducibility and weakens the interpretation of reported multimodal gains. Rather than proposing a new detection architecture, this work introduces an auditable evaluation framework based on three explicit stream contracts: RGB Exhaustive (Rx), Depth Exhaustive (Dx), and RGB Co-observable (Rc). The framework formalizes modality selection, co-observability rules, dataset cardinality, and strict identity-level pairing under the official train/validation/test partitions of a public RGB–depth AAV benchmark. YOLOv8n serves as the reference detector, while YOLO11n parity experiments, cardinality-controlled RGB baselines, zero-shot cross-dataset evaluation, and additional backbone controls provide complementary analyses without changing the underlying protocol. Pairing statistics reveal that only 44.1% of training depth frames and 10.1% of test depth frames satisfy the released identity-level strict pairing rule (shared numeric identity and equal image dimensions), showing that protocol-level constraints substantially influence the sample space available for multimodal evaluation. Rc is an RGB-only detector trained on the depth-availability-selected RGB subset; it is not a jointly trained RGB–depth model. Under a matched training budget and multiple cardinality-controlled Rk subsets, uniform RGB subsampling outperforms co-observable selection on the official RGB test, whereas Rc leads on the n=135 paired test population—showing that identity-level pairing changes evaluation populations rather than uniformly improving detection. All primary tabulated metrics are taken from frozen training exports to support auditability and to avoid post hoc checkpoint cherry-picking. The proposed framework makes co-observability rules, annotation ownership, and dataset composition explicit experimental variables for multimodal AAV detection.