Page 100 - Read Online
P. 100
Page 254 Shi et al. Art Int Surg 2024;4:247-57 https://dx.doi.org/10.20517/ais.2024.17
Figure 3. Qualitative comparison of predicted depth map on the SCARED dataset between our method with SOTA depth estimation
methods. For visualization and evaluation purposes, the ground-truth depth values are scaled. SCARED: Stereo correspondence and
reconstruction of endoscopic data; SOTA: state-of-the-art.
Figure 5 illustrates an example of 3D surface reconstruction using the predicted depth of our method over
Lite-Mono and AF-SfMLearner. When analyzing the reconstruction in Figure 5, it is evident that visual
distortions or errors exist at the edges of objects, making it challenging to accurately represent spatial
relationships between objects. This difficulty is particularly pronounced at the edges or transitions between
different regions, but in the depth map visualization shown in Figure 3, these interpolation artifacts may not
be visible. Our model results provided a greatly superior surface prediction outcome, which is crucial
because the reconstruction outcomes reflect a major true objective of monocular depth estimation.
