Effective object detection using Tensorflow facilitated YOLOv3 model

Abstract

Object detection forms an important area of research where the efforts are still being put forth to improve the accuracy of detection. Several approaches have been made which also include R-CNN and DNN. Whereas they have rendered the detection of object to be more cumbersome as each component has to be trained separately. This also poses a challenge while optimizing and hence takes more time for single detection. Whereas, in the present study, an attempt is made to effectively locate and detect the objects as a single regression problem, thereby reducing the time for image detection. The technique incorporated is YOLOv3 with significant assistance from tensorflow. An in – depth understanding is achieved and the performance of the model is assessed. Python is used for processing the images and to prepare the YOLOv3 model. The training of the model is facilitated by Pascal Visual Object Classes(VOC) dataset, which comprises of nearly 11000 images for testing, 5717 images for training and 5823 images for validation, which are found to be quite sufficient to attain good detection accuracy. The training is carried out initially for a learning rate of 1e – 4 with first 20 epochs, followed by 30 epochs for a learning rate of 1e – 6. The model is trained with the help of available Common Objects in Context (COCO) weights. The precision of each object detected is evaluated and the mean Average Precision (mAP) is found to be around 79%.