TinyROIFusion is a lightweight framework designed for window-based detection of tiny and multi-scale objects in high-resolution images. By leveraging object tracking-based Region of Interest (ROI) proposals, it enhances detection accuracy while maintaining computational efficiency. Detecting small and tiny objects in high-resolution images presents significant challenges due to their minimal pixel representation and the complexity of backgrounds. TinyROIFusion addresses these challenges by integrating object tracking mechanisms into segmentation-based ROI estimation. This approach generates precise ROI proposals, improving the performance of lightweight detection models.
This repository provides the implementation of the methods described in the following publication:
Kos, A., Majek, K. and Belter, D., Enhanced Lightweight Detection of Small and Tiny Objects in High-Resolution Images Using Object Tracking-Based Region of Interest Proposal. Available at SSRN 4951026.
We provide a Docker environment that manages all dependencies, downloads pretrained models, and sets up datasets automatically.
git clone https://github.com/deepdrivepl/TinyROIFusion.git && cd TinyROIFusion
./build_and_run.shTo run inference on the SeaDronesSee and DroneCrowd datasets, execute ./exps.sh. This script runs the inference process and saves predictions in COCO format in the results/SeaDronesSee and results/DroneCrowd directories. It also generates visualizations of ROIs, detection windows, and detected objects.
python inference_sequence.py --out_dir results/SeaDronesSee \
--ds SeaDronesSee --split val \
--mode 'roi_track' \
--merge \
--obs_iou_th 0.1 \
--debug --vis_conf_th 0.1 --show_labelpython inference_sequence.py --out_dir results/DroneCrowd \
--ds DroneCrowd --flist data/DroneCrowd/test_minus_val.txt --name test_minus_val \
--mode 'roi_track' \
--merge \
--obs_iou_th 0.7 --second_nms \
--debug --vis_conf_th 0.1This project builds upon existing research in object detection, tracking, and region proposal methods.