Beyond Appearance: Dual-Graph Object Encoding With Learnable Graph Structure

Object encoding is essential for enabling robots to efficiently perform tasks such as recognition and autonomous exploration. Existing object encoding approaches typically rely on appearance-based representations, leading to poor performance in environments with multiple visually similar objects, which can compromise downstream task accuracy. Inspired by human perception of objects through both appearance and structure, we introduce DGOE, a novel object encoding method, which adopts a dual-graph embedding scheme. Rather than treating object representation as a single-level appearance encoding problem, DGOE explicitly models object discriminability through a dual-level structural formulation, decomposing it into intrinsic inner-object structure and cross-object relational context. DGOE leverages graph structures as the carrier to integrate object intrinsic appearance and cross-object spatial features. This unified framework leads to more distinctive and reliable object representations. Additionally, we use multi-head self-attention to learn cross-object graph structures for extracting cross-object spatial features. Experimental evaluations on multiple publicly available datasets demonstrate the superior performance of our method in object-level matching tasks, underscoring its effectiveness and robustness.