Improving the Accuracy of Object Detection Utilizing Camouflage Models

In aerial reconnaissance, it is necessary to detect military targets. However, due to their scarcity, it brings the problem of few shot object detection. But training only with the physical camouflage models will cause the object detection model degradation. In order to enhance the ability of object detection in this scenario, a plug-and-play post-processing method is proposed. First, we train a binary classification model in the simulation domain to achieve the correct classification of the foreground and background. Lower the threshold of any object detection method, the classifier will obtain the probability that each bounding box belongs to the foreground object. Then, through the pseudo attention mechanism, the probabilities of the detection model and the classifier are further calculated, sorted and filtered, and finally the score of each bounding box is output. Experiments on HIMARS images show that the proposed method has a certain improvement on the two few shot detection methods, TFA and DeFRCN. Especially when applied to the cross-domain detection method CDFSOD, it can increase the average accuracy of zero shot detection to 0.343.