Figure 2. (a) For the COCO dataset, all the images are resized such that the short edge has 800 pixels while the long edge has less than
1333 pixels. For the ImageNet dataset, all the images are resized to 224×224 to calculate the statistics. The scale of object is represented
as the square root of the area. We discuss the typical training input size for ImageNet classification and COCO detection tasks. (b) All the
scores are predicted on COCO minival dataset by the RetinaNet detector with FPN and ResNet 50 backbone, which is trained with 10%
COCO data. The score distance is the absolute difference between the predictions of the image in different sizes. The Y-axis is the average
numberofanchorsperimage. (c)Wepredictpseudo-labelontherestofCOCOtrainingdatawithaconvergedFaster-RCNNdetector(with
FPNandResNet50backbone), trainedwith10%COCOdata. Thelowaveragerecallandprecisionshowthathardpseudo-labelincurmore
noise with False Negative samples.