An Improved Object Detection Algorithm Based on CenterNet

CenterNet (Object as Points) is a recently popular single-stage anchor free object detection algorithm. First, this article adds a new attention module in which the mean and maximum values of the channel are used, it also introduces variance information to better express the feature distribution of each layer from different aspects. Second, we improved the loss function and replaced the original L1 loss function with the IoU loss function, so that the loss function of the borders and the center point regression has the scale invariance, which is consistent with the final standard IoU judgment, and will not increase the model inference time. Finally, multi-scale images are used for training. By enriching the number of samples of different sizes, the problem of model overfitting due to less data is effectively reduced. This paper compares the improved CenterNet algorithm with the original algorithm on the Pascal VOC data set for training and testing. The final model increases the mAP by 3.5% under the unaugmented test and 3.03% mAP for using horizontal flipping augmented test. In terms of time overhead, there is only a small expenditure, it can still maintain the detection speed of 12 ms per frame.