Frontiers of Computer Vision by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9789811548185
Publisher: Springer Singapore

In the proposed model, the initial values of all layers of the backbone structure use pre-trained models, ResNet50, ResNet101, and ResNet152, which are trained with ImageNet [12]. The datasets used in this paper are PASCAL VOC 2007, VOC 2012 and COCO2017. The VOC 2007 and 2012 datasets contain for 21 classes, including objects in 20 kinds and backgrounds. The COCO2017 is a Common Objects in Context that is a large-scale object detection, segmentation, and captioning dataset. It has 80 objects and contains images, bounding boxes and labels for the 2017 version. The Table 1 shows the number of images and object ground truth as follows. Table 2 shows the results with the various training parameters. In the original paper [3], Faster R-CNN shows the original experiment results are 69.9% mAP for VGG16’s backbone network and voc2007 dataset, 73.2% mAP for voc2007 and 2012, and 78.8% mAP for additional coco dataset. Original ResNet50, 101, 152 models are trained on the voc2007 dataset in Table 2. For each model, mAP(%) results of 70.22, 75.66 and 78.57 are shown. The results are used to compare with the proposed models. The proposed model shows the results according to anchor size, backbone and dilated rate. The experimental results are generated using ResNet101, and 152 models, and shows 74.91, 76.69% mAP when dilated rate = 6, 12, 18. The following results are 61.65, 71.15, 74.08% mAP with anchor size = 64, 128, 256, 512 considering small objects, which is 2.53.8% lower than previous results. When extracting features from RPN, the performance decreases with the large value of dilated rates. The reason is that it is a difficult to extract the feature. Because it looks around the wide receptive field which is relatively unnecessary in the existing feature map. The final results conduct the change in dilated rate. In the case of backbone network in ResNet152 to achieve 79.56% mAP for dilated rate = 3, 6, 9, 79.83% mAP for dilated rate = 2, 4, 8, and 79.12% mAP for dilated rate = 1, 2, 3. In the combination of dilated feature maps, dilated rate = 2, 4, 8 is 0.27% and 0.71% better than dilated rate = 3, 6, 9 and 1, 2, 3, respectively. There are different results from voc2007 dataset. The voc2012 dataset is the double number of dataset than voc2007. One of reason that the performance is a different for approximately 13% mAP. The size of the input feature map is reduced to 1/16 of the original image, and the dilated rate is considered the wide receptive field. However the feature extraction of dilated rate = 2, 4, 8 is 3.14% mAP higher than that of dilated rate = 6, 12, 18. In the case of coco dataset, the proposed models conduct with two dilated rate. The evaluation of coco2017 are shown in Table 3. The results are generated by training the proposed model for ResNet152. AP @ [0.5: 0.95] is the result of calculating the average AP for ten IOU values with IOU threshold values between 0.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.