Imagine a sleek, modern dining table (Gvenet) surrounded by wildly colorful, mismatched chairs (Alice). Imagine a crisp white bedroom (Gvenet) with a pile of bright, quirky pillows and a stack of books on the floor (Alice).
: The graph structure naturally captures long-range dependencies and occlusion relationships, outperforming grid-based CNNs, especially for small objects (e.g., 12% better on “fork” in COCO). gvenet and alice
Given an input image ( I ), we use a superpixel algorithm (SLIC) to generate ( N ) nodes ( V = v_1, \dots, v_N ). Each node ( v_i ) is associated with a feature vector ( f_i \in \mathbbR^d ) extracted via a lightweight CNN backbone. Edges ( E ) connect nodes that are spatially adjacent or have high feature similarity (k-NN in feature space, ( k=5 )). Imagine a sleek, modern dining table (Gvenet) surrounded