DMP: Temporal Context in Computer Vision Detection Model

This dataset contains evaluation results, model artefacts and analysis outputs from experiments investigating the impact of temporal context on robustness to visual perturbations in video object detection. Models based on YOLOV-Swin and YOLOX architectures were trained using PyTorch on the VisDrone dataset (public, CC BY-NC-SA 3.0) and evaluated under both clean conditions and simulated noise levels ranging from 10 % to 60 %. In addition, the XS-VID dataset (MIT license) was used to assess model performance in scenarios dominated by extremely small objects, enabling targeted evaluation of temporal modelling under challenging visual conditions.

The dataset includes performance metrics (mAP, mAR, inference time) in CSV format, visual performance plots, qualitative prediction examples, trained model checkpoints and summary reports. All scripts, model configurations and environment specifications required to reproduce or extend the experiments are provided in the accompanying code repository:
https://github.com/mozi30/TemporalAttentionPlayground

Noise-based robustness results were generated programmatically using performance degradation models, rather than full retraining on corrupted data. Further evaluation using additional architectures such as MSDA and TransVOD is planned as future work.

This dataset is intended to support reproducible experimentation and comparative benchmarking in drone-based perception research, temporal attention modelling and robust computer vision development.