Object detection has achieved a tremendous advancement based on feature-based learning in the vision space, while little work has focused on reasoning in the perception space like humans. One of the greatest challenges lies in that it is difficult to build a connectivity model in the topological space for relational reasoning, since the current network is better at modeling the distribution of structured data. To settle this issue, we introduce a novel graph modeling mechanism with class-based graph representation, which contributes to modeling the high-order topology structure that maps the data distribution to make the detection models have better relational reasoning ability. In this mechanism, we propose three learning subtasks, i.e., vision-to-perception embedding, perception reasoning graph representation, and perception-to-vision modeling. The mechanism based on such subtasks effectively maintains the independence of the original detection network and the proposed mechanism-based model, thus it can be well integrated with existing detection models without additional modification. The experimental results demonstrate the feasibility and effectiveness of our proposed mechanism, and the new state-of-the-art performance can be achieved on the public challenging datasets for object detection.
展开▼
