Object detection methods based on convolutional neural networks have achieved great success in many computer vision tasks. However, most existing object detection methods are designed based on the hypothesis of closed category set, and thus, can only detect objects of predetermined categories appearing in the training set, while ignoring objects of unknown categories as background. Therefore, this thesis focuses on open set object detection, aiming to detect objects of unknown categories and mark them as unknown categories while detecting the objects of predetermined categories. Several methods are proposed and their effectiveness is verified in experiments. The main contributions are as follows:
1. An open set object detection method based on the rejection of abnormal objects is proposed. This method combines the existing detection algorithm with the open set recognition algorithm, and uses the one-versus-all criterion for training, breaking the closed set assumption of the classification layer in the conventional detection algorithm, so as to enable the discovery of unknown category objects. Satisfactory results have been achieved on the task of detecting unknown insects in natural scenes.
2. A one-stage open-set object detection method based on prototype learning is proposed. This method uses the prototype based classifier to form a compact intraclass distribution and a scattered inter-class distribution in the feature space. This can effectively improve the unknown object detection performance while maintaining the detection ability of known objects. Experiments on PASCAL VOC datasets show that
the proposed method can effectively identify unknown class objects while ensuring the detection performance of known class objects.
3. A training strategy based on saliency guidance is proposed. It uses the saliency information in the image as a clue to guide the detector to respond positively to the unlabeled object regions in the image, thereby overcoming the adverse effect of unlabeled foreground objects in the training image on the discovery of new categories. At the same time, the algorithm can use more data to fully train the model and effectively improve the ability to discover unknown targets. Experiments on the PASCAL VOC dataset show that our proposed strategy significantly improves the performance of the CenterNet-based open-set object detector on object discovery tasks.
修改评论