Fast weakly supervised object detection via image pixel gradient

  • Rui LI, Si’qiang LIU


In recent years, the attention of weakly supervised object detection has greatly increased. This is due to the advantage that weakly supervised object detection only needs image-level annotation information during the training process, further reducing the cost of detection. Most of the existing weakly-supervised object detection algorithms use region proposal generation network to generate region proposal to determine the object category and location, and then generate pseudo-labels and input it into the fully-supervised network to predict the object category and location. This leads to a large amount of low-level information being ignored and falling into a local optimum due to the influence of noise. To solve the above problems, this paper proposes a weakly supervised detection model similar to the SSD framework, which uses the image pixel gradient map (IPG) to locate potential objects in the object image, and uses object instance mining for each potential object in order to improve the positioning accuracy of the model. Use the mining result as a pseudo ground-truth to fine-tune in the model to obtain the final detection result. Experiments on the PASACL VOC 2007 and 2012 datasets prove that the accuracy of this model is higher than that of the existing weakly supervised object detection model, and the detection speed of the model without any object proposal is significantly improved.