Résumé: In the problem of few-shot object detection, class prototype knowledge in previous works is not be fully refined and utilized due to lack of instances. We noticed that the application of the output features of the RoI pooling layer has a great influence on the grasp of the prototype features, which motivates us to focus on how to reuse them. Therefore, we propose a multiple knowledge embedding network, which gets improvement in three places in the fine-tuning stage. Introducing attention mechanism to strengthen feature extraction, Up-CoTNet is used to replace the 3×3 convolution in Resnet101. Feature enhancement module is added to enforce the common object reasoning for the output of RoI pooling layer. Then, we propose a contrastive learning branch to grasp the information encoded between different feature regions. Experiments on PASCAL VOC and MS COCO datasets show that our model significantly raises the performance by 1.3% (+0.5AP) in average compared with previous methods.

Langue: Anglais