Results

Figure 1. Comparisons of affordance detection. The first two rows show that our method outperform the original AffordanceNet[1]and can avoid the noisy fragments in affordance maps. The third row shows that our hierarchical AffordanceNet may lose some fine-grained information as it involves the high-level affordance classification.

Figure 2. Samples of baseline comparison. We view those point clouds from different orientation to clearly compare the reconstruction of the "contain" regions. As is shown in this figure, our method can achieve stable reconstruction of transparent objects, especially the inner the surfaces of cups.

Table I. Baseline comparisons on different evaluation regions