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.

Grasp detection plays a critical role for robot manipulation. Mainstream pixel-wise grasp detection networks with encoder-decoder structure receive much attention due to good accuracy and efficiency. However, they usually transmit the high-level feature in the encoder to the decoder, and low-level features are neglected. It is noted that low-level features contain abundant detail information, and how to fully exploit low-level features remains unsolved. Meanwhile, the channel information in high-level feature is also not well mined. Inevitably, the performance of grasp detection is degraded. To solve these problems, we propose a grasp detection network with hierarchical multi-scale feature fusion and inverted shuffle residual. Both low-level and high-level features in the encoder are firstly fused by the designed skip connections with attention module, and the fused information is then propagated to corresponding layers of the decoder for in-depth feature fusion. Such a hierarchical fusion guarantees the quality of grasp prediction. Furthermore, an inverted shuffle residual module is created, where the high-level feature from encoder is split in channel and the resultant split features are processed in their respective branches. By such differentiation processing, more high-dimensional channel information is kept, which enhances the representation ability of the network. Besides, an information enhancement module is added before the encoder to reinforce input information. The proposed method attains 98.9% and 97.8% in image-wise and object-wise accuracy on the Cornell grasping dataset, respectively, and the experimental results verify the effectiveness of the method.