Performance Comparison of YOLOv7 and YOLOv8 Using the YCB Datasets YCB-M and YCB-Video

EasyChair Preprint 13111, version 1

Abstract

In this paper, the two YOLO frameworks, YOLOv7 and YOLOv8, are compared using the two labeled YCB datasets, YCB-M and YCB-Video. In addition, a custom test dataset is created in the robotics context to observe how the two YOLO frameworks perform on dissimilar data (compared to the training data). The performance is measured by considering the mean average precision (mAP), average detection time, and resource consumption. Furthermore, the impact of different amounts of training data on performance is observed. For com- parability, a training and validation pipeline is established that every trained model undergoes. We were able to show that both frameworks perform very well and sim- ilarly on similar datas (test data from the two datasets YCB-M and YCB-Video) and on dissimilar data (the custom-created test dataset), YOLOv7 significantly outperforms YOLOv8 by 22% mAP.

Keyphrases: MS COCO, Service Robotics, YCB Dataset, YOLO, benchmark, map, object detection

@booklet{EasyChair:13111,
  author    = {Samuel Hafner and Markus Schneider and Benjamin Stähle},
  title     = {Performance Comparison of YOLOv7 and YOLOv8 Using the YCB Datasets YCB-M and YCB-Video},
  howpublished = {EasyChair Preprint 13111},
  year      = {EasyChair, 2024}}