Fig 15. F1-Confidence curve for increased bias Fig16. Precision- Recall curve for increased bias
6. CONCLUSION & FUTURE WORK
Object detection includes basic tasks including object categorization, location, and segmentation. Object identification and related activities are divided into two groups: single stage and two stage. In this paper, we explored the single stage detection using YOLO V5 for object detection using xView Dataset. The paper examined single stage object detectors, particularly YOLO v 5, their architectural developments, underlying pre-trained CNN architectures. It also includes the several elements, optimizations, and tweaking of YOLO v5 hyper parameters, as well as all underlying combinations, in order to get the optimal model for improved identification of multi class objects.
Table 1 represents the summary of the experiments conducted with various combinations of the hyper-parameters tuned to evaluate all the different application scenarios. The best value for the parameters is mentioned in th table for the model. These experiments resulted in deciding the best combination of the hyper parameters and finalsing the model for the best detection using the xView dataset. Similar experiments can be conducted with other versions of YOLO and tested on variety of data sets. By tuning the hyper parameters the model is fully exercised to give better results and suitably the best model with correct combination of hyper parameters and optimisers is achieved.
YOLOs are being used extensively in many real-time object tracking applications. Convolutional neural network advancements should be extensively tested in order to enhance single stage object detectors. Depending on the type of applications and underlying datasets, various algorithms may be sutilized to further optimise the process.