Performance evaluation of fine-tuned faster R-CNN on specific MS COCO objects

Abstract

Fine-tuning of a model is a method that is most often required to cater the users' explicit requirements. But the question remains whether the model is accurate enough to be used for a certain application. This paper strives to present the metrics used for performance evaluation of a Convolutional Neural Network (CNN) model. The evaluation is based on the training process which provides us with intermediate models after every 1000 iterations. While 1000 iterations is not substantial enough over the range of 490k iterations, the groups are sized with 100k iterations each. Now, the intention was to compare the recorded metrics to evaluate the model in terms of accuracy. The training model used the set of specific categories chosen from the Microsoft Common Objects in Context (MS COCO) dataset, while allowing users to use their own externally available images to test the model's accuracy. Our trained model ensured that all the objects are detected that are present in the image to depict the effect of precision.