Yolov3 Bottle Detector

A Yolov3-based bottle brand detector, which is trained from a custom dataset with four brands of mineral water bottles.

Custom Detection Dataset:

This water bottle detection dataset consists of 4870 images of four brands of mineral water bottles (i.e. Aquafina, Deer, Kirkland and Nestle). Images were collected by the turtlebot 2 robot and smart phone in four different environments: lobby, laboratory, corridor, and living room.

Size

• 4000 training images
• 870 validation images

Bottle Classes

• Aquafina
• Deer
• Kirkland
• Nestle

Format

• PASCAL VOC
• Darknet

Download link

• https://zenodo.org/record/7065974
• DOI: 10.5281/zenodo.7065974

Dataset Folder Structure

• Annotations: contains the xml label files in PASCAL VOC format
• ImageSets: contains the training index files
• JPEGImages: contains the image data in jpg format
• Labels: contains the txt label files in Darknet format

Sample labeled Images

YOLOV3 Tiny Bottle Detector:

This bottle detector is a pretrained yolov3-tiny model fine-tuned by our custom bottle dataset shown above.

Network Configure File

• ./cfg/yolov3-tiny-sphd.cfg

Two pretrained models

• ./weights/yolov3_tiny_sphd_25000_paper.weights: specially used in our SPHD filter paper "The Semantic PHD Filter for Multi-class TargetTracking: From Theory to Practice"
• ./weights/yolov3_tiny_30000_general.weights: general purpose (recommend)

Requirements

• install darknet package: https://github.com/pjreddie/darknet

Video Demo

• ./demo/bottle_detection_demo.mp4

Citation

@article{chen2022semantic,
  title={The semantic PHD filter for multi-class target tracking: From theory to practice},
  author={Chen, Jun and Xie, Zhanteng and Dames, Philip},
  journal={Robotics and Autonomous Systems},
  volume={149},
  pages={103947},
  year={2022},
  publisher={Elsevier}
}

@article{xie2022dataset,
  title={Experimental Datasets and Processing Codes for the Semantic PHD Filter},
  author={Xie, Zhanteng and Chen, Jun and Dames, Philip},
  year={2022},
}