While using the moge-2 model for depth inference, the output depth tensor frequently contains Inf values, which leads to invalid statistics and breaks downstream visualization.
Specifically:
depth_map contains infinite.
depth_normalized = self._normalize_depth_for_visualization(depth_estimation.depth_map)
plt.imshow(depth_normalized, cmap='plasma')
plt.colorbar()
When visualizing the depth map with matplotlib, the result appears entirely white.
_save_results in depth_estimation.py
depth_fixed = depth_estimation.depth_map.copy()
depth_fixed[np.isinf(depth_fixed)] = np.nan
valid_max = np.nanmax(depth_fixed)
depth_map = np.nan_to_num(depth_fixed, nan=valid_max)
# Save depth map as image
depth_normalized = self._normalize_depth_for_visualization(depth_map)
before
after
While using the moge-2 model for depth inference, the output depth tensor frequently contains Inf values, which leads to invalid statistics and breaks downstream visualization.
Specifically:
depth_map contains infinite.
When visualizing the depth map with matplotlib, the result appears entirely white.
_save_resultsindepth_estimation.pybefore
after
