Few-shot object detection receives much attention with the ability to detect novel class objects using limited annotated data. The transfer learning-based solution becomes popular due to its simple training with good accuracy, however, it is still challenging to enrich the feature diversity during the training process. And fine-grained features are also insufficient for novel class detection. To deal with the problems, this paper proposes a novel few-shot object detection method based on dual-domain feature fusion and patch-level attention. Upon original base domain, an elementary domain with more category-agnostic features is superposed to construct a two-stream backbone, which benefits to enrich the feature diversity. To better integrate various features, a dual-domain feature fusion is designed, where the feature pairs with the same size are complementarily fused to extract more discriminative features. Moreover, a patch-wise feature refinement termed as patch-level attention is presented to mine internal relations among the patches, which enhances the adaptability to novel classes. In addition, a weighted classification loss is given to assist the fine-tuning of the classifier by combining extra features from FPN of the base training model. In this way, the few-shot detection quality to novel class objects is improved. Experiments on PASCAL VOC and MS COCO datasets verify the effectiveness of the method.

The ability to recognize novel objects with a few visual samples is critical in the robotic applications. Existing methods mainly concern the recognition of inter-category objects, however, the object recognition from different sub-classes within the same category remains challenging due to their similar appearances. In this paper, we propose a key-part attention retrieval solution to distinguish novel objects of different sub-classes according to a few samples without re-training. Especially, an object encoder, including convolutional neural network with attention and key-part aggregation, is designed to generate object attention map and extract the object-level embedding, where object attention map from the middle stage of the backbone is used to guide the key-part aggregation. Besides, to overcome the non-differentiability drawback of key-part attention, the object encoder is trained in a two-step scheme, and a more stable object-level embedding is obtained. On this basis, the potential objects are located from a scene image by mining connected domains of the attention map. By matching the embedding of each potential object and embeddings from support data, the recognition of the potential objects is achieved. The effectiveness of the proposed method is verified by experiments.