| Object detection is one of the fundamental tasks in the field of computer vision.It has a wide range of applications such as security monitoring,unmanned vehicle,human-computer interaction,and so on.In this paper,we study and analyze the advantages and disadvantages of the existing deep neural network based object detection algorithms.In order to deal with the disadvantages,we propose two object detection algorithms.The performance of object detection is highly reliant on the confidence scores,the bounding boxes,and the instance segmentation masks.In this paper,we focus on improving the confidence scores to achieve better performance for object detection.We revisit and discuss the importance of the confidence scores in object detection and propose a new framework to improve the confidence scores for object detection in three aspects:(1)we train an IoU(Intersection-over-Union)estimator to predict the IoU for each region of interest box with the ground-truth box and incorporate the predicted IoU into the final confidence scores to distinguish the quality of the bounding boxes;(2)we decouple the classification branch into the foreground/background classification and the category classification to avoid each category competing with a large number of negative samples;(3)we propose the interleaved execution for the box branch and the confidence scores related branches to alleviate misalignment of features.By taking advantage of the improved confidence scores,the experiments show that the proposed algorithm can consistently improve the performance of baseline detectors on the challenging MSCOCO dataset.Object detection for aerial images is becoming an active topic in computer vision so that we also design a detection method for aerial images.It is a very challenging task due to many factors such as highly complex background,arbitrary object orientations,small objects,and high input resolution,etc.In order to improve the training efficiency and the detection accuracy,in this paper,we develop a suitable training and inference mechanism for high resolution aerial images and we design appropriate sampling method and data augmentation strategy to address the issues of the larger input resolution,small objects and most areas as the background in aerial object detection.Instead of estimating the orientations of objects using direct regressions like that in previous methods,we propose to predict the rotated bounding boxes by leveraging a segmentation task,which is easier to train and yields more accurate detection results.Extensive experiments have been conducted to verify the effectiveness of the algorithm.The proposed method achieves state of the art performance on the DOTA dataset which is a very challenging aerial images dataset. |
